Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 *
21 * Fixes:
22 * Alan Cox : Numerous verify_area() calls
23 * Alan Cox : Set the ACK bit on a reset
24 * Alan Cox : Stopped it crashing if it closed while
25 * sk->inuse=1 and was trying to connect
26 * (tcp_err()).
27 * Alan Cox : All icmp error handling was broken
28 * pointers passed where wrong and the
29 * socket was looked up backwards. Nobody
30 * tested any icmp error code obviously.
31 * Alan Cox : tcp_err() now handled properly. It
32 * wakes people on errors. poll
33 * behaves and the icmp error race
34 * has gone by moving it into sock.c
35 * Alan Cox : tcp_send_reset() fixed to work for
36 * everything not just packets for
37 * unknown sockets.
38 * Alan Cox : tcp option processing.
39 * Alan Cox : Reset tweaked (still not 100%) [Had
40 * syn rule wrong]
41 * Herp Rosmanith : More reset fixes
42 * Alan Cox : No longer acks invalid rst frames.
43 * Acking any kind of RST is right out.
44 * Alan Cox : Sets an ignore me flag on an rst
45 * receive otherwise odd bits of prattle
46 * escape still
47 * Alan Cox : Fixed another acking RST frame bug.
48 * Should stop LAN workplace lockups.
49 * Alan Cox : Some tidyups using the new skb list
50 * facilities
51 * Alan Cox : sk->keepopen now seems to work
52 * Alan Cox : Pulls options out correctly on accepts
53 * Alan Cox : Fixed assorted sk->rqueue->next errors
54 * Alan Cox : PSH doesn't end a TCP read. Switched a
55 * bit to skb ops.
56 * Alan Cox : Tidied tcp_data to avoid a potential
57 * nasty.
58 * Alan Cox : Added some better commenting, as the
59 * tcp is hard to follow
60 * Alan Cox : Removed incorrect check for 20 * psh
61 * Michael O'Reilly : ack < copied bug fix.
62 * Johannes Stille : Misc tcp fixes (not all in yet).
63 * Alan Cox : FIN with no memory -> CRASH
64 * Alan Cox : Added socket option proto entries.
65 * Also added awareness of them to accept.
66 * Alan Cox : Added TCP options (SOL_TCP)
67 * Alan Cox : Switched wakeup calls to callbacks,
68 * so the kernel can layer network
69 * sockets.
70 * Alan Cox : Use ip_tos/ip_ttl settings.
71 * Alan Cox : Handle FIN (more) properly (we hope).
72 * Alan Cox : RST frames sent on unsynchronised
73 * state ack error.
74 * Alan Cox : Put in missing check for SYN bit.
75 * Alan Cox : Added tcp_select_window() aka NET2E
76 * window non shrink trick.
77 * Alan Cox : Added a couple of small NET2E timer
78 * fixes
79 * Charles Hedrick : TCP fixes
80 * Toomas Tamm : TCP window fixes
81 * Alan Cox : Small URG fix to rlogin ^C ack fight
82 * Charles Hedrick : Rewrote most of it to actually work
83 * Linus : Rewrote tcp_read() and URG handling
84 * completely
85 * Gerhard Koerting: Fixed some missing timer handling
86 * Matthew Dillon : Reworked TCP machine states as per RFC
87 * Gerhard Koerting: PC/TCP workarounds
88 * Adam Caldwell : Assorted timer/timing errors
89 * Matthew Dillon : Fixed another RST bug
90 * Alan Cox : Move to kernel side addressing changes.
91 * Alan Cox : Beginning work on TCP fastpathing
92 * (not yet usable)
93 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
94 * Alan Cox : TCP fast path debugging
95 * Alan Cox : Window clamping
96 * Michael Riepe : Bug in tcp_check()
97 * Matt Dillon : More TCP improvements and RST bug fixes
98 * Matt Dillon : Yet more small nasties remove from the
99 * TCP code (Be very nice to this man if
100 * tcp finally works 100%) 8)
101 * Alan Cox : BSD accept semantics.
102 * Alan Cox : Reset on closedown bug.
103 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
104 * Michael Pall : Handle poll() after URG properly in
105 * all cases.
106 * Michael Pall : Undo the last fix in tcp_read_urg()
107 * (multi URG PUSH broke rlogin).
108 * Michael Pall : Fix the multi URG PUSH problem in
109 * tcp_readable(), poll() after URG
110 * works now.
111 * Michael Pall : recv(...,MSG_OOB) never blocks in the
112 * BSD api.
113 * Alan Cox : Changed the semantics of sk->socket to
114 * fix a race and a signal problem with
115 * accept() and async I/O.
116 * Alan Cox : Relaxed the rules on tcp_sendto().
117 * Yury Shevchuk : Really fixed accept() blocking problem.
118 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
119 * clients/servers which listen in on
120 * fixed ports.
121 * Alan Cox : Cleaned the above up and shrank it to
122 * a sensible code size.
123 * Alan Cox : Self connect lockup fix.
124 * Alan Cox : No connect to multicast.
125 * Ross Biro : Close unaccepted children on master
126 * socket close.
127 * Alan Cox : Reset tracing code.
128 * Alan Cox : Spurious resets on shutdown.
129 * Alan Cox : Giant 15 minute/60 second timer error
130 * Alan Cox : Small whoops in polling before an
131 * accept.
132 * Alan Cox : Kept the state trace facility since
133 * it's handy for debugging.
134 * Alan Cox : More reset handler fixes.
135 * Alan Cox : Started rewriting the code based on
136 * the RFC's for other useful protocol
137 * references see: Comer, KA9Q NOS, and
138 * for a reference on the difference
139 * between specifications and how BSD
140 * works see the 4.4lite source.
141 * A.N.Kuznetsov : Don't time wait on completion of tidy
142 * close.
143 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
144 * Linus Torvalds : Fixed BSD port reuse to work first syn
145 * Alan Cox : Reimplemented timers as per the RFC
146 * and using multiple timers for sanity.
147 * Alan Cox : Small bug fixes, and a lot of new
148 * comments.
149 * Alan Cox : Fixed dual reader crash by locking
150 * the buffers (much like datagram.c)
151 * Alan Cox : Fixed stuck sockets in probe. A probe
152 * now gets fed up of retrying without
153 * (even a no space) answer.
154 * Alan Cox : Extracted closing code better
155 * Alan Cox : Fixed the closing state machine to
156 * resemble the RFC.
157 * Alan Cox : More 'per spec' fixes.
158 * Jorge Cwik : Even faster checksumming.
159 * Alan Cox : tcp_data() doesn't ack illegal PSH
160 * only frames. At least one pc tcp stack
161 * generates them.
162 * Alan Cox : Cache last socket.
163 * Alan Cox : Per route irtt.
164 * Matt Day : poll()->select() match BSD precisely on error
165 * Alan Cox : New buffers
166 * Marc Tamsky : Various sk->prot->retransmits and
167 * sk->retransmits misupdating fixed.
168 * Fixed tcp_write_timeout: stuck close,
169 * and TCP syn retries gets used now.
170 * Mark Yarvis : In tcp_read_wakeup(), don't send an
171 * ack if state is TCP_CLOSED.
172 * Alan Cox : Look up device on a retransmit - routes may
173 * change. Doesn't yet cope with MSS shrink right
174 * but it's a start!
175 * Marc Tamsky : Closing in closing fixes.
176 * Mike Shaver : RFC1122 verifications.
177 * Alan Cox : rcv_saddr errors.
178 * Alan Cox : Block double connect().
179 * Alan Cox : Small hooks for enSKIP.
180 * Alexey Kuznetsov: Path MTU discovery.
181 * Alan Cox : Support soft errors.
182 * Alan Cox : Fix MTU discovery pathological case
183 * when the remote claims no mtu!
184 * Marc Tamsky : TCP_CLOSE fix.
185 * Colin (G3TNE) : Send a reset on syn ack replies in
186 * window but wrong (fixes NT lpd problems)
187 * Pedro Roque : Better TCP window handling, delayed ack.
188 * Joerg Reuter : No modification of locked buffers in
189 * tcp_do_retransmit()
190 * Eric Schenk : Changed receiver side silly window
191 * avoidance algorithm to BSD style
192 * algorithm. This doubles throughput
193 * against machines running Solaris,
194 * and seems to result in general
195 * improvement.
196 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
197 * Willy Konynenberg : Transparent proxying support.
198 * Mike McLagan : Routing by source
199 * Keith Owens : Do proper merging with partial SKB's in
200 * tcp_do_sendmsg to avoid burstiness.
201 * Eric Schenk : Fix fast close down bug with
202 * shutdown() followed by close().
203 * Andi Kleen : Make poll agree with SIGIO
204 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
205 * lingertime == 0 (RFC 793 ABORT Call)
206 * Hirokazu Takahashi : Use copy_from_user() instead of
207 * csum_and_copy_from_user() if possible.
208 *
209 * Description of States:
210 *
211 * TCP_SYN_SENT sent a connection request, waiting for ack
212 *
213 * TCP_SYN_RECV received a connection request, sent ack,
214 * waiting for final ack in three-way handshake.
215 *
216 * TCP_ESTABLISHED connection established
217 *
218 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
219 * transmission of remaining buffered data
220 *
221 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
222 * to shutdown
223 *
224 * TCP_CLOSING both sides have shutdown but we still have
225 * data we have to finish sending
226 *
227 * TCP_TIME_WAIT timeout to catch resent junk before entering
228 * closed, can only be entered from FIN_WAIT2
229 * or CLOSING. Required because the other end
230 * may not have gotten our last ACK causing it
231 * to retransmit the data packet (which we ignore)
232 *
233 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
234 * us to finish writing our data and to shutdown
235 * (we have to close() to move on to LAST_ACK)
236 *
237 * TCP_LAST_ACK out side has shutdown after remote has
238 * shutdown. There may still be data in our
239 * buffer that we have to finish sending
240 *
241 * TCP_CLOSE socket is finished
242 */
243
244#define pr_fmt(fmt) "TCP: " fmt
245
246#include <crypto/hash.h>
247#include <linux/kernel.h>
248#include <linux/module.h>
249#include <linux/types.h>
250#include <linux/fcntl.h>
251#include <linux/poll.h>
252#include <linux/inet_diag.h>
253#include <linux/init.h>
254#include <linux/fs.h>
255#include <linux/skbuff.h>
256#include <linux/scatterlist.h>
257#include <linux/splice.h>
258#include <linux/net.h>
259#include <linux/socket.h>
260#include <linux/random.h>
261#include <linux/memblock.h>
262#include <linux/highmem.h>
263#include <linux/swap.h>
264#include <linux/cache.h>
265#include <linux/err.h>
266#include <linux/time.h>
267#include <linux/slab.h>
268#include <linux/errqueue.h>
269#include <linux/static_key.h>
270#include <linux/btf.h>
271
272#include <net/icmp.h>
273#include <net/inet_common.h>
274#include <net/tcp.h>
275#include <net/mptcp.h>
276#include <net/xfrm.h>
277#include <net/ip.h>
278#include <net/sock.h>
279
280#include <linux/uaccess.h>
281#include <asm/ioctls.h>
282#include <net/busy_poll.h>
283
284/* Track pending CMSGs. */
285enum {
286 TCP_CMSG_INQ = 1,
287 TCP_CMSG_TS = 2
288};
289
290struct percpu_counter tcp_orphan_count;
291EXPORT_SYMBOL_GPL(tcp_orphan_count);
292
293long sysctl_tcp_mem[3] __read_mostly;
294EXPORT_SYMBOL(sysctl_tcp_mem);
295
296atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
297EXPORT_SYMBOL(tcp_memory_allocated);
298
299#if IS_ENABLED(CONFIG_SMC)
300DEFINE_STATIC_KEY_FALSE(tcp_have_smc);
301EXPORT_SYMBOL(tcp_have_smc);
302#endif
303
304/*
305 * Current number of TCP sockets.
306 */
307struct percpu_counter tcp_sockets_allocated;
308EXPORT_SYMBOL(tcp_sockets_allocated);
309
310/*
311 * TCP splice context
312 */
313struct tcp_splice_state {
314 struct pipe_inode_info *pipe;
315 size_t len;
316 unsigned int flags;
317};
318
319/*
320 * Pressure flag: try to collapse.
321 * Technical note: it is used by multiple contexts non atomically.
322 * All the __sk_mem_schedule() is of this nature: accounting
323 * is strict, actions are advisory and have some latency.
324 */
325unsigned long tcp_memory_pressure __read_mostly;
326EXPORT_SYMBOL_GPL(tcp_memory_pressure);
327
328DEFINE_STATIC_KEY_FALSE(tcp_rx_skb_cache_key);
329EXPORT_SYMBOL(tcp_rx_skb_cache_key);
330
331DEFINE_STATIC_KEY_FALSE(tcp_tx_skb_cache_key);
332
333void tcp_enter_memory_pressure(struct sock *sk)
334{
335 unsigned long val;
336
337 if (READ_ONCE(tcp_memory_pressure))
338 return;
339 val = jiffies;
340
341 if (!val)
342 val--;
343 if (!cmpxchg(&tcp_memory_pressure, 0, val))
344 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
345}
346EXPORT_SYMBOL_GPL(tcp_enter_memory_pressure);
347
348void tcp_leave_memory_pressure(struct sock *sk)
349{
350 unsigned long val;
351
352 if (!READ_ONCE(tcp_memory_pressure))
353 return;
354 val = xchg(&tcp_memory_pressure, 0);
355 if (val)
356 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURESCHRONO,
357 jiffies_to_msecs(jiffies - val));
358}
359EXPORT_SYMBOL_GPL(tcp_leave_memory_pressure);
360
361/* Convert seconds to retransmits based on initial and max timeout */
362static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
363{
364 u8 res = 0;
365
366 if (seconds > 0) {
367 int period = timeout;
368
369 res = 1;
370 while (seconds > period && res < 255) {
371 res++;
372 timeout <<= 1;
373 if (timeout > rto_max)
374 timeout = rto_max;
375 period += timeout;
376 }
377 }
378 return res;
379}
380
381/* Convert retransmits to seconds based on initial and max timeout */
382static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
383{
384 int period = 0;
385
386 if (retrans > 0) {
387 period = timeout;
388 while (--retrans) {
389 timeout <<= 1;
390 if (timeout > rto_max)
391 timeout = rto_max;
392 period += timeout;
393 }
394 }
395 return period;
396}
397
398static u64 tcp_compute_delivery_rate(const struct tcp_sock *tp)
399{
400 u32 rate = READ_ONCE(tp->rate_delivered);
401 u32 intv = READ_ONCE(tp->rate_interval_us);
402 u64 rate64 = 0;
403
404 if (rate && intv) {
405 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
406 do_div(rate64, intv);
407 }
408 return rate64;
409}
410
411/* Address-family independent initialization for a tcp_sock.
412 *
413 * NOTE: A lot of things set to zero explicitly by call to
414 * sk_alloc() so need not be done here.
415 */
416void tcp_init_sock(struct sock *sk)
417{
418 struct inet_connection_sock *icsk = inet_csk(sk);
419 struct tcp_sock *tp = tcp_sk(sk);
420
421 tp->out_of_order_queue = RB_ROOT;
422 sk->tcp_rtx_queue = RB_ROOT;
423 tcp_init_xmit_timers(sk);
424 INIT_LIST_HEAD(&tp->tsq_node);
425 INIT_LIST_HEAD(&tp->tsorted_sent_queue);
426
427 icsk->icsk_rto = TCP_TIMEOUT_INIT;
428 icsk->icsk_rto_min = TCP_RTO_MIN;
429 icsk->icsk_delack_max = TCP_DELACK_MAX;
430 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
431 minmax_reset(&tp->rtt_min, tcp_jiffies32, ~0U);
432
433 /* So many TCP implementations out there (incorrectly) count the
434 * initial SYN frame in their delayed-ACK and congestion control
435 * algorithms that we must have the following bandaid to talk
436 * efficiently to them. -DaveM
437 */
438 tp->snd_cwnd = TCP_INIT_CWND;
439
440 /* There's a bubble in the pipe until at least the first ACK. */
441 tp->app_limited = ~0U;
442
443 /* See draft-stevens-tcpca-spec-01 for discussion of the
444 * initialization of these values.
445 */
446 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
447 tp->snd_cwnd_clamp = ~0;
448 tp->mss_cache = TCP_MSS_DEFAULT;
449
450 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
451 tcp_assign_congestion_control(sk);
452
453 tp->tsoffset = 0;
454 tp->rack.reo_wnd_steps = 1;
455
456 sk->sk_write_space = sk_stream_write_space;
457 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
458
459 icsk->icsk_sync_mss = tcp_sync_mss;
460
461 WRITE_ONCE(sk->sk_sndbuf, sock_net(sk)->ipv4.sysctl_tcp_wmem[1]);
462 WRITE_ONCE(sk->sk_rcvbuf, sock_net(sk)->ipv4.sysctl_tcp_rmem[1]);
463
464 sk_sockets_allocated_inc(sk);
465 sk->sk_route_forced_caps = NETIF_F_GSO;
466}
467EXPORT_SYMBOL(tcp_init_sock);
468
469static void tcp_tx_timestamp(struct sock *sk, u16 tsflags)
470{
471 struct sk_buff *skb = tcp_write_queue_tail(sk);
472
473 if (tsflags && skb) {
474 struct skb_shared_info *shinfo = skb_shinfo(skb);
475 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
476
477 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
478 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
479 tcb->txstamp_ack = 1;
480 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
481 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
482 }
483}
484
485static bool tcp_stream_is_readable(struct sock *sk, int target)
486{
487 if (tcp_epollin_ready(sk, target))
488 return true;
489
490 if (sk->sk_prot->stream_memory_read)
491 return sk->sk_prot->stream_memory_read(sk);
492 return false;
493}
494
495/*
496 * Wait for a TCP event.
497 *
498 * Note that we don't need to lock the socket, as the upper poll layers
499 * take care of normal races (between the test and the event) and we don't
500 * go look at any of the socket buffers directly.
501 */
502__poll_t tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
503{
504 __poll_t mask;
505 struct sock *sk = sock->sk;
506 const struct tcp_sock *tp = tcp_sk(sk);
507 int state;
508
509 sock_poll_wait(file, sock, wait);
510
511 state = inet_sk_state_load(sk);
512 if (state == TCP_LISTEN)
513 return inet_csk_listen_poll(sk);
514
515 /* Socket is not locked. We are protected from async events
516 * by poll logic and correct handling of state changes
517 * made by other threads is impossible in any case.
518 */
519
520 mask = 0;
521
522 /*
523 * EPOLLHUP is certainly not done right. But poll() doesn't
524 * have a notion of HUP in just one direction, and for a
525 * socket the read side is more interesting.
526 *
527 * Some poll() documentation says that EPOLLHUP is incompatible
528 * with the EPOLLOUT/POLLWR flags, so somebody should check this
529 * all. But careful, it tends to be safer to return too many
530 * bits than too few, and you can easily break real applications
531 * if you don't tell them that something has hung up!
532 *
533 * Check-me.
534 *
535 * Check number 1. EPOLLHUP is _UNMASKABLE_ event (see UNIX98 and
536 * our fs/select.c). It means that after we received EOF,
537 * poll always returns immediately, making impossible poll() on write()
538 * in state CLOSE_WAIT. One solution is evident --- to set EPOLLHUP
539 * if and only if shutdown has been made in both directions.
540 * Actually, it is interesting to look how Solaris and DUX
541 * solve this dilemma. I would prefer, if EPOLLHUP were maskable,
542 * then we could set it on SND_SHUTDOWN. BTW examples given
543 * in Stevens' books assume exactly this behaviour, it explains
544 * why EPOLLHUP is incompatible with EPOLLOUT. --ANK
545 *
546 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
547 * blocking on fresh not-connected or disconnected socket. --ANK
548 */
549 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
550 mask |= EPOLLHUP;
551 if (sk->sk_shutdown & RCV_SHUTDOWN)
552 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
553
554 /* Connected or passive Fast Open socket? */
555 if (state != TCP_SYN_SENT &&
556 (state != TCP_SYN_RECV || rcu_access_pointer(tp->fastopen_rsk))) {
557 int target = sock_rcvlowat(sk, 0, INT_MAX);
558
559 if (READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq) &&
560 !sock_flag(sk, SOCK_URGINLINE) &&
561 tp->urg_data)
562 target++;
563
564 if (tcp_stream_is_readable(sk, target))
565 mask |= EPOLLIN | EPOLLRDNORM;
566
567 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
568 if (__sk_stream_is_writeable(sk, 1)) {
569 mask |= EPOLLOUT | EPOLLWRNORM;
570 } else { /* send SIGIO later */
571 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
572 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
573
574 /* Race breaker. If space is freed after
575 * wspace test but before the flags are set,
576 * IO signal will be lost. Memory barrier
577 * pairs with the input side.
578 */
579 smp_mb__after_atomic();
580 if (__sk_stream_is_writeable(sk, 1))
581 mask |= EPOLLOUT | EPOLLWRNORM;
582 }
583 } else
584 mask |= EPOLLOUT | EPOLLWRNORM;
585
586 if (tp->urg_data & TCP_URG_VALID)
587 mask |= EPOLLPRI;
588 } else if (state == TCP_SYN_SENT && inet_sk(sk)->defer_connect) {
589 /* Active TCP fastopen socket with defer_connect
590 * Return EPOLLOUT so application can call write()
591 * in order for kernel to generate SYN+data
592 */
593 mask |= EPOLLOUT | EPOLLWRNORM;
594 }
595 /* This barrier is coupled with smp_wmb() in tcp_reset() */
596 smp_rmb();
597 if (sk->sk_err || !skb_queue_empty_lockless(&sk->sk_error_queue))
598 mask |= EPOLLERR;
599
600 return mask;
601}
602EXPORT_SYMBOL(tcp_poll);
603
604int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
605{
606 struct tcp_sock *tp = tcp_sk(sk);
607 int answ;
608 bool slow;
609
610 switch (cmd) {
611 case SIOCINQ:
612 if (sk->sk_state == TCP_LISTEN)
613 return -EINVAL;
614
615 slow = lock_sock_fast(sk);
616 answ = tcp_inq(sk);
617 unlock_sock_fast(sk, slow);
618 break;
619 case SIOCATMARK:
620 answ = tp->urg_data &&
621 READ_ONCE(tp->urg_seq) == READ_ONCE(tp->copied_seq);
622 break;
623 case SIOCOUTQ:
624 if (sk->sk_state == TCP_LISTEN)
625 return -EINVAL;
626
627 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
628 answ = 0;
629 else
630 answ = READ_ONCE(tp->write_seq) - tp->snd_una;
631 break;
632 case SIOCOUTQNSD:
633 if (sk->sk_state == TCP_LISTEN)
634 return -EINVAL;
635
636 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
637 answ = 0;
638 else
639 answ = READ_ONCE(tp->write_seq) -
640 READ_ONCE(tp->snd_nxt);
641 break;
642 default:
643 return -ENOIOCTLCMD;
644 }
645
646 return put_user(answ, (int __user *)arg);
647}
648EXPORT_SYMBOL(tcp_ioctl);
649
650static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
651{
652 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
653 tp->pushed_seq = tp->write_seq;
654}
655
656static inline bool forced_push(const struct tcp_sock *tp)
657{
658 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
659}
660
661static void skb_entail(struct sock *sk, struct sk_buff *skb)
662{
663 struct tcp_sock *tp = tcp_sk(sk);
664 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
665
666 skb->csum = 0;
667 tcb->seq = tcb->end_seq = tp->write_seq;
668 tcb->tcp_flags = TCPHDR_ACK;
669 tcb->sacked = 0;
670 __skb_header_release(skb);
671 tcp_add_write_queue_tail(sk, skb);
672 sk_wmem_queued_add(sk, skb->truesize);
673 sk_mem_charge(sk, skb->truesize);
674 if (tp->nonagle & TCP_NAGLE_PUSH)
675 tp->nonagle &= ~TCP_NAGLE_PUSH;
676
677 tcp_slow_start_after_idle_check(sk);
678}
679
680static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
681{
682 if (flags & MSG_OOB)
683 tp->snd_up = tp->write_seq;
684}
685
686/* If a not yet filled skb is pushed, do not send it if
687 * we have data packets in Qdisc or NIC queues :
688 * Because TX completion will happen shortly, it gives a chance
689 * to coalesce future sendmsg() payload into this skb, without
690 * need for a timer, and with no latency trade off.
691 * As packets containing data payload have a bigger truesize
692 * than pure acks (dataless) packets, the last checks prevent
693 * autocorking if we only have an ACK in Qdisc/NIC queues,
694 * or if TX completion was delayed after we processed ACK packet.
695 */
696static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
697 int size_goal)
698{
699 return skb->len < size_goal &&
700 sock_net(sk)->ipv4.sysctl_tcp_autocorking &&
701 !tcp_rtx_queue_empty(sk) &&
702 refcount_read(&sk->sk_wmem_alloc) > skb->truesize;
703}
704
705void tcp_push(struct sock *sk, int flags, int mss_now,
706 int nonagle, int size_goal)
707{
708 struct tcp_sock *tp = tcp_sk(sk);
709 struct sk_buff *skb;
710
711 skb = tcp_write_queue_tail(sk);
712 if (!skb)
713 return;
714 if (!(flags & MSG_MORE) || forced_push(tp))
715 tcp_mark_push(tp, skb);
716
717 tcp_mark_urg(tp, flags);
718
719 if (tcp_should_autocork(sk, skb, size_goal)) {
720
721 /* avoid atomic op if TSQ_THROTTLED bit is already set */
722 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
724 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
725 }
726 /* It is possible TX completion already happened
727 * before we set TSQ_THROTTLED.
728 */
729 if (refcount_read(&sk->sk_wmem_alloc) > skb->truesize)
730 return;
731 }
732
733 if (flags & MSG_MORE)
734 nonagle = TCP_NAGLE_CORK;
735
736 __tcp_push_pending_frames(sk, mss_now, nonagle);
737}
738
739static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
740 unsigned int offset, size_t len)
741{
742 struct tcp_splice_state *tss = rd_desc->arg.data;
743 int ret;
744
745 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
746 min(rd_desc->count, len), tss->flags);
747 if (ret > 0)
748 rd_desc->count -= ret;
749 return ret;
750}
751
752static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
753{
754 /* Store TCP splice context information in read_descriptor_t. */
755 read_descriptor_t rd_desc = {
756 .arg.data = tss,
757 .count = tss->len,
758 };
759
760 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
761}
762
763/**
764 * tcp_splice_read - splice data from TCP socket to a pipe
765 * @sock: socket to splice from
766 * @ppos: position (not valid)
767 * @pipe: pipe to splice to
768 * @len: number of bytes to splice
769 * @flags: splice modifier flags
770 *
771 * Description:
772 * Will read pages from given socket and fill them into a pipe.
773 *
774 **/
775ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
776 struct pipe_inode_info *pipe, size_t len,
777 unsigned int flags)
778{
779 struct sock *sk = sock->sk;
780 struct tcp_splice_state tss = {
781 .pipe = pipe,
782 .len = len,
783 .flags = flags,
784 };
785 long timeo;
786 ssize_t spliced;
787 int ret;
788
789 sock_rps_record_flow(sk);
790 /*
791 * We can't seek on a socket input
792 */
793 if (unlikely(*ppos))
794 return -ESPIPE;
795
796 ret = spliced = 0;
797
798 lock_sock(sk);
799
800 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
801 while (tss.len) {
802 ret = __tcp_splice_read(sk, &tss);
803 if (ret < 0)
804 break;
805 else if (!ret) {
806 if (spliced)
807 break;
808 if (sock_flag(sk, SOCK_DONE))
809 break;
810 if (sk->sk_err) {
811 ret = sock_error(sk);
812 break;
813 }
814 if (sk->sk_shutdown & RCV_SHUTDOWN)
815 break;
816 if (sk->sk_state == TCP_CLOSE) {
817 /*
818 * This occurs when user tries to read
819 * from never connected socket.
820 */
821 ret = -ENOTCONN;
822 break;
823 }
824 if (!timeo) {
825 ret = -EAGAIN;
826 break;
827 }
828 /* if __tcp_splice_read() got nothing while we have
829 * an skb in receive queue, we do not want to loop.
830 * This might happen with URG data.
831 */
832 if (!skb_queue_empty(&sk->sk_receive_queue))
833 break;
834 sk_wait_data(sk, &timeo, NULL);
835 if (signal_pending(current)) {
836 ret = sock_intr_errno(timeo);
837 break;
838 }
839 continue;
840 }
841 tss.len -= ret;
842 spliced += ret;
843
844 if (!timeo)
845 break;
846 release_sock(sk);
847 lock_sock(sk);
848
849 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
850 (sk->sk_shutdown & RCV_SHUTDOWN) ||
851 signal_pending(current))
852 break;
853 }
854
855 release_sock(sk);
856
857 if (spliced)
858 return spliced;
859
860 return ret;
861}
862EXPORT_SYMBOL(tcp_splice_read);
863
864struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
865 bool force_schedule)
866{
867 struct sk_buff *skb;
868
869 if (likely(!size)) {
870 skb = sk->sk_tx_skb_cache;
871 if (skb) {
872 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
873 sk->sk_tx_skb_cache = NULL;
874 pskb_trim(skb, 0);
875 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
876 skb_shinfo(skb)->tx_flags = 0;
877 memset(TCP_SKB_CB(skb), 0, sizeof(struct tcp_skb_cb));
878 return skb;
879 }
880 }
881 /* The TCP header must be at least 32-bit aligned. */
882 size = ALIGN(size, 4);
883
884 if (unlikely(tcp_under_memory_pressure(sk)))
885 sk_mem_reclaim_partial(sk);
886
887 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
888 if (likely(skb)) {
889 bool mem_scheduled;
890
891 if (force_schedule) {
892 mem_scheduled = true;
893 sk_forced_mem_schedule(sk, skb->truesize);
894 } else {
895 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
896 }
897 if (likely(mem_scheduled)) {
898 skb_reserve(skb, sk->sk_prot->max_header);
899 /*
900 * Make sure that we have exactly size bytes
901 * available to the caller, no more, no less.
902 */
903 skb->reserved_tailroom = skb->end - skb->tail - size;
904 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
905 return skb;
906 }
907 __kfree_skb(skb);
908 } else {
909 sk->sk_prot->enter_memory_pressure(sk);
910 sk_stream_moderate_sndbuf(sk);
911 }
912 return NULL;
913}
914
915static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
916 int large_allowed)
917{
918 struct tcp_sock *tp = tcp_sk(sk);
919 u32 new_size_goal, size_goal;
920
921 if (!large_allowed)
922 return mss_now;
923
924 /* Note : tcp_tso_autosize() will eventually split this later */
925 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
926 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
927
928 /* We try hard to avoid divides here */
929 size_goal = tp->gso_segs * mss_now;
930 if (unlikely(new_size_goal < size_goal ||
931 new_size_goal >= size_goal + mss_now)) {
932 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
933 sk->sk_gso_max_segs);
934 size_goal = tp->gso_segs * mss_now;
935 }
936
937 return max(size_goal, mss_now);
938}
939
940int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
941{
942 int mss_now;
943
944 mss_now = tcp_current_mss(sk);
945 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
946
947 return mss_now;
948}
949
950/* In some cases, both sendpage() and sendmsg() could have added
951 * an skb to the write queue, but failed adding payload on it.
952 * We need to remove it to consume less memory, but more
953 * importantly be able to generate EPOLLOUT for Edge Trigger epoll()
954 * users.
955 */
956void tcp_remove_empty_skb(struct sock *sk, struct sk_buff *skb)
957{
958 if (skb && !skb->len) {
959 tcp_unlink_write_queue(skb, sk);
960 if (tcp_write_queue_empty(sk))
961 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
962 sk_wmem_free_skb(sk, skb);
963 }
964}
965
966struct sk_buff *tcp_build_frag(struct sock *sk, int size_goal, int flags,
967 struct page *page, int offset, size_t *size)
968{
969 struct sk_buff *skb = tcp_write_queue_tail(sk);
970 struct tcp_sock *tp = tcp_sk(sk);
971 bool can_coalesce;
972 int copy, i;
973
974 if (!skb || (copy = size_goal - skb->len) <= 0 ||
975 !tcp_skb_can_collapse_to(skb)) {
976new_segment:
977 if (!sk_stream_memory_free(sk))
978 return NULL;
979
980 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
981 tcp_rtx_and_write_queues_empty(sk));
982 if (!skb)
983 return NULL;
984
985#ifdef CONFIG_TLS_DEVICE
986 skb->decrypted = !!(flags & MSG_SENDPAGE_DECRYPTED);
987#endif
988 skb_entail(sk, skb);
989 copy = size_goal;
990 }
991
992 if (copy > *size)
993 copy = *size;
994
995 i = skb_shinfo(skb)->nr_frags;
996 can_coalesce = skb_can_coalesce(skb, i, page, offset);
997 if (!can_coalesce && i >= sysctl_max_skb_frags) {
998 tcp_mark_push(tp, skb);
999 goto new_segment;
1000 }
1001 if (!sk_wmem_schedule(sk, copy))
1002 return NULL;
1003
1004 if (can_coalesce) {
1005 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1006 } else {
1007 get_page(page);
1008 skb_fill_page_desc(skb, i, page, offset, copy);
1009 }
1010
1011 if (!(flags & MSG_NO_SHARED_FRAGS))
1012 skb_shinfo(skb)->flags |= SKBFL_SHARED_FRAG;
1013
1014 skb->len += copy;
1015 skb->data_len += copy;
1016 skb->truesize += copy;
1017 sk_wmem_queued_add(sk, copy);
1018 sk_mem_charge(sk, copy);
1019 skb->ip_summed = CHECKSUM_PARTIAL;
1020 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1021 TCP_SKB_CB(skb)->end_seq += copy;
1022 tcp_skb_pcount_set(skb, 0);
1023
1024 *size = copy;
1025 return skb;
1026}
1027
1028ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
1029 size_t size, int flags)
1030{
1031 struct tcp_sock *tp = tcp_sk(sk);
1032 int mss_now, size_goal;
1033 int err;
1034 ssize_t copied;
1035 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1036
1037 if (IS_ENABLED(CONFIG_DEBUG_VM) &&
1038 WARN_ONCE(!sendpage_ok(page),
1039 "page must not be a Slab one and have page_count > 0"))
1040 return -EINVAL;
1041
1042 /* Wait for a connection to finish. One exception is TCP Fast Open
1043 * (passive side) where data is allowed to be sent before a connection
1044 * is fully established.
1045 */
1046 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1047 !tcp_passive_fastopen(sk)) {
1048 err = sk_stream_wait_connect(sk, &timeo);
1049 if (err != 0)
1050 goto out_err;
1051 }
1052
1053 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1054
1055 mss_now = tcp_send_mss(sk, &size_goal, flags);
1056 copied = 0;
1057
1058 err = -EPIPE;
1059 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1060 goto out_err;
1061
1062 while (size > 0) {
1063 struct sk_buff *skb;
1064 size_t copy = size;
1065
1066 skb = tcp_build_frag(sk, size_goal, flags, page, offset, ©);
1067 if (!skb)
1068 goto wait_for_space;
1069
1070 if (!copied)
1071 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1072
1073 copied += copy;
1074 offset += copy;
1075 size -= copy;
1076 if (!size)
1077 goto out;
1078
1079 if (skb->len < size_goal || (flags & MSG_OOB))
1080 continue;
1081
1082 if (forced_push(tp)) {
1083 tcp_mark_push(tp, skb);
1084 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1085 } else if (skb == tcp_send_head(sk))
1086 tcp_push_one(sk, mss_now);
1087 continue;
1088
1089wait_for_space:
1090 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1091 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1092 TCP_NAGLE_PUSH, size_goal);
1093
1094 err = sk_stream_wait_memory(sk, &timeo);
1095 if (err != 0)
1096 goto do_error;
1097
1098 mss_now = tcp_send_mss(sk, &size_goal, flags);
1099 }
1100
1101out:
1102 if (copied) {
1103 tcp_tx_timestamp(sk, sk->sk_tsflags);
1104 if (!(flags & MSG_SENDPAGE_NOTLAST))
1105 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1106 }
1107 return copied;
1108
1109do_error:
1110 tcp_remove_empty_skb(sk, tcp_write_queue_tail(sk));
1111 if (copied)
1112 goto out;
1113out_err:
1114 /* make sure we wake any epoll edge trigger waiter */
1115 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1116 sk->sk_write_space(sk);
1117 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1118 }
1119 return sk_stream_error(sk, flags, err);
1120}
1121EXPORT_SYMBOL_GPL(do_tcp_sendpages);
1122
1123int tcp_sendpage_locked(struct sock *sk, struct page *page, int offset,
1124 size_t size, int flags)
1125{
1126 if (!(sk->sk_route_caps & NETIF_F_SG))
1127 return sock_no_sendpage_locked(sk, page, offset, size, flags);
1128
1129 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1130
1131 return do_tcp_sendpages(sk, page, offset, size, flags);
1132}
1133EXPORT_SYMBOL_GPL(tcp_sendpage_locked);
1134
1135int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1136 size_t size, int flags)
1137{
1138 int ret;
1139
1140 lock_sock(sk);
1141 ret = tcp_sendpage_locked(sk, page, offset, size, flags);
1142 release_sock(sk);
1143
1144 return ret;
1145}
1146EXPORT_SYMBOL(tcp_sendpage);
1147
1148void tcp_free_fastopen_req(struct tcp_sock *tp)
1149{
1150 if (tp->fastopen_req) {
1151 kfree(tp->fastopen_req);
1152 tp->fastopen_req = NULL;
1153 }
1154}
1155
1156static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1157 int *copied, size_t size,
1158 struct ubuf_info *uarg)
1159{
1160 struct tcp_sock *tp = tcp_sk(sk);
1161 struct inet_sock *inet = inet_sk(sk);
1162 struct sockaddr *uaddr = msg->msg_name;
1163 int err, flags;
1164
1165 if (!(sock_net(sk)->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) ||
1166 (uaddr && msg->msg_namelen >= sizeof(uaddr->sa_family) &&
1167 uaddr->sa_family == AF_UNSPEC))
1168 return -EOPNOTSUPP;
1169 if (tp->fastopen_req)
1170 return -EALREADY; /* Another Fast Open is in progress */
1171
1172 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1173 sk->sk_allocation);
1174 if (unlikely(!tp->fastopen_req))
1175 return -ENOBUFS;
1176 tp->fastopen_req->data = msg;
1177 tp->fastopen_req->size = size;
1178 tp->fastopen_req->uarg = uarg;
1179
1180 if (inet->defer_connect) {
1181 err = tcp_connect(sk);
1182 /* Same failure procedure as in tcp_v4/6_connect */
1183 if (err) {
1184 tcp_set_state(sk, TCP_CLOSE);
1185 inet->inet_dport = 0;
1186 sk->sk_route_caps = 0;
1187 }
1188 }
1189 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1190 err = __inet_stream_connect(sk->sk_socket, uaddr,
1191 msg->msg_namelen, flags, 1);
1192 /* fastopen_req could already be freed in __inet_stream_connect
1193 * if the connection times out or gets rst
1194 */
1195 if (tp->fastopen_req) {
1196 *copied = tp->fastopen_req->copied;
1197 tcp_free_fastopen_req(tp);
1198 inet->defer_connect = 0;
1199 }
1200 return err;
1201}
1202
1203int tcp_sendmsg_locked(struct sock *sk, struct msghdr *msg, size_t size)
1204{
1205 struct tcp_sock *tp = tcp_sk(sk);
1206 struct ubuf_info *uarg = NULL;
1207 struct sk_buff *skb;
1208 struct sockcm_cookie sockc;
1209 int flags, err, copied = 0;
1210 int mss_now = 0, size_goal, copied_syn = 0;
1211 int process_backlog = 0;
1212 bool zc = false;
1213 long timeo;
1214
1215 flags = msg->msg_flags;
1216
1217 if (flags & MSG_ZEROCOPY && size && sock_flag(sk, SOCK_ZEROCOPY)) {
1218 skb = tcp_write_queue_tail(sk);
1219 uarg = msg_zerocopy_realloc(sk, size, skb_zcopy(skb));
1220 if (!uarg) {
1221 err = -ENOBUFS;
1222 goto out_err;
1223 }
1224
1225 zc = sk->sk_route_caps & NETIF_F_SG;
1226 if (!zc)
1227 uarg->zerocopy = 0;
1228 }
1229
1230 if (unlikely(flags & MSG_FASTOPEN || inet_sk(sk)->defer_connect) &&
1231 !tp->repair) {
1232 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size, uarg);
1233 if (err == -EINPROGRESS && copied_syn > 0)
1234 goto out;
1235 else if (err)
1236 goto out_err;
1237 }
1238
1239 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1240
1241 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1242
1243 /* Wait for a connection to finish. One exception is TCP Fast Open
1244 * (passive side) where data is allowed to be sent before a connection
1245 * is fully established.
1246 */
1247 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1248 !tcp_passive_fastopen(sk)) {
1249 err = sk_stream_wait_connect(sk, &timeo);
1250 if (err != 0)
1251 goto do_error;
1252 }
1253
1254 if (unlikely(tp->repair)) {
1255 if (tp->repair_queue == TCP_RECV_QUEUE) {
1256 copied = tcp_send_rcvq(sk, msg, size);
1257 goto out_nopush;
1258 }
1259
1260 err = -EINVAL;
1261 if (tp->repair_queue == TCP_NO_QUEUE)
1262 goto out_err;
1263
1264 /* 'common' sending to sendq */
1265 }
1266
1267 sockcm_init(&sockc, sk);
1268 if (msg->msg_controllen) {
1269 err = sock_cmsg_send(sk, msg, &sockc);
1270 if (unlikely(err)) {
1271 err = -EINVAL;
1272 goto out_err;
1273 }
1274 }
1275
1276 /* This should be in poll */
1277 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1278
1279 /* Ok commence sending. */
1280 copied = 0;
1281
1282restart:
1283 mss_now = tcp_send_mss(sk, &size_goal, flags);
1284
1285 err = -EPIPE;
1286 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1287 goto do_error;
1288
1289 while (msg_data_left(msg)) {
1290 int copy = 0;
1291
1292 skb = tcp_write_queue_tail(sk);
1293 if (skb)
1294 copy = size_goal - skb->len;
1295
1296 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1297 bool first_skb;
1298
1299new_segment:
1300 if (!sk_stream_memory_free(sk))
1301 goto wait_for_space;
1302
1303 if (unlikely(process_backlog >= 16)) {
1304 process_backlog = 0;
1305 if (sk_flush_backlog(sk))
1306 goto restart;
1307 }
1308 first_skb = tcp_rtx_and_write_queues_empty(sk);
1309 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
1310 first_skb);
1311 if (!skb)
1312 goto wait_for_space;
1313
1314 process_backlog++;
1315 skb->ip_summed = CHECKSUM_PARTIAL;
1316
1317 skb_entail(sk, skb);
1318 copy = size_goal;
1319
1320 /* All packets are restored as if they have
1321 * already been sent. skb_mstamp_ns isn't set to
1322 * avoid wrong rtt estimation.
1323 */
1324 if (tp->repair)
1325 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1326 }
1327
1328 /* Try to append data to the end of skb. */
1329 if (copy > msg_data_left(msg))
1330 copy = msg_data_left(msg);
1331
1332 /* Where to copy to? */
1333 if (skb_availroom(skb) > 0 && !zc) {
1334 /* We have some space in skb head. Superb! */
1335 copy = min_t(int, copy, skb_availroom(skb));
1336 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1337 if (err)
1338 goto do_fault;
1339 } else if (!zc) {
1340 bool merge = true;
1341 int i = skb_shinfo(skb)->nr_frags;
1342 struct page_frag *pfrag = sk_page_frag(sk);
1343
1344 if (!sk_page_frag_refill(sk, pfrag))
1345 goto wait_for_space;
1346
1347 if (!skb_can_coalesce(skb, i, pfrag->page,
1348 pfrag->offset)) {
1349 if (i >= sysctl_max_skb_frags) {
1350 tcp_mark_push(tp, skb);
1351 goto new_segment;
1352 }
1353 merge = false;
1354 }
1355
1356 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1357
1358 if (!sk_wmem_schedule(sk, copy))
1359 goto wait_for_space;
1360
1361 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1362 pfrag->page,
1363 pfrag->offset,
1364 copy);
1365 if (err)
1366 goto do_error;
1367
1368 /* Update the skb. */
1369 if (merge) {
1370 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1371 } else {
1372 skb_fill_page_desc(skb, i, pfrag->page,
1373 pfrag->offset, copy);
1374 page_ref_inc(pfrag->page);
1375 }
1376 pfrag->offset += copy;
1377 } else {
1378 if (!sk_wmem_schedule(sk, copy))
1379 goto wait_for_space;
1380
1381 err = skb_zerocopy_iter_stream(sk, skb, msg, copy, uarg);
1382 if (err == -EMSGSIZE || err == -EEXIST) {
1383 tcp_mark_push(tp, skb);
1384 goto new_segment;
1385 }
1386 if (err < 0)
1387 goto do_error;
1388 copy = err;
1389 }
1390
1391 if (!copied)
1392 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1393
1394 WRITE_ONCE(tp->write_seq, tp->write_seq + copy);
1395 TCP_SKB_CB(skb)->end_seq += copy;
1396 tcp_skb_pcount_set(skb, 0);
1397
1398 copied += copy;
1399 if (!msg_data_left(msg)) {
1400 if (unlikely(flags & MSG_EOR))
1401 TCP_SKB_CB(skb)->eor = 1;
1402 goto out;
1403 }
1404
1405 if (skb->len < size_goal || (flags & MSG_OOB) || unlikely(tp->repair))
1406 continue;
1407
1408 if (forced_push(tp)) {
1409 tcp_mark_push(tp, skb);
1410 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1411 } else if (skb == tcp_send_head(sk))
1412 tcp_push_one(sk, mss_now);
1413 continue;
1414
1415wait_for_space:
1416 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1417 if (copied)
1418 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1419 TCP_NAGLE_PUSH, size_goal);
1420
1421 err = sk_stream_wait_memory(sk, &timeo);
1422 if (err != 0)
1423 goto do_error;
1424
1425 mss_now = tcp_send_mss(sk, &size_goal, flags);
1426 }
1427
1428out:
1429 if (copied) {
1430 tcp_tx_timestamp(sk, sockc.tsflags);
1431 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1432 }
1433out_nopush:
1434 net_zcopy_put(uarg);
1435 return copied + copied_syn;
1436
1437do_error:
1438 skb = tcp_write_queue_tail(sk);
1439do_fault:
1440 tcp_remove_empty_skb(sk, skb);
1441
1442 if (copied + copied_syn)
1443 goto out;
1444out_err:
1445 net_zcopy_put_abort(uarg, true);
1446 err = sk_stream_error(sk, flags, err);
1447 /* make sure we wake any epoll edge trigger waiter */
1448 if (unlikely(tcp_rtx_and_write_queues_empty(sk) && err == -EAGAIN)) {
1449 sk->sk_write_space(sk);
1450 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1451 }
1452 return err;
1453}
1454EXPORT_SYMBOL_GPL(tcp_sendmsg_locked);
1455
1456int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1457{
1458 int ret;
1459
1460 lock_sock(sk);
1461 ret = tcp_sendmsg_locked(sk, msg, size);
1462 release_sock(sk);
1463
1464 return ret;
1465}
1466EXPORT_SYMBOL(tcp_sendmsg);
1467
1468/*
1469 * Handle reading urgent data. BSD has very simple semantics for
1470 * this, no blocking and very strange errors 8)
1471 */
1472
1473static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1474{
1475 struct tcp_sock *tp = tcp_sk(sk);
1476
1477 /* No URG data to read. */
1478 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1479 tp->urg_data == TCP_URG_READ)
1480 return -EINVAL; /* Yes this is right ! */
1481
1482 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1483 return -ENOTCONN;
1484
1485 if (tp->urg_data & TCP_URG_VALID) {
1486 int err = 0;
1487 char c = tp->urg_data;
1488
1489 if (!(flags & MSG_PEEK))
1490 tp->urg_data = TCP_URG_READ;
1491
1492 /* Read urgent data. */
1493 msg->msg_flags |= MSG_OOB;
1494
1495 if (len > 0) {
1496 if (!(flags & MSG_TRUNC))
1497 err = memcpy_to_msg(msg, &c, 1);
1498 len = 1;
1499 } else
1500 msg->msg_flags |= MSG_TRUNC;
1501
1502 return err ? -EFAULT : len;
1503 }
1504
1505 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1506 return 0;
1507
1508 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1509 * the available implementations agree in this case:
1510 * this call should never block, independent of the
1511 * blocking state of the socket.
1512 * Mike <pall@rz.uni-karlsruhe.de>
1513 */
1514 return -EAGAIN;
1515}
1516
1517static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1518{
1519 struct sk_buff *skb;
1520 int copied = 0, err = 0;
1521
1522 /* XXX -- need to support SO_PEEK_OFF */
1523
1524 skb_rbtree_walk(skb, &sk->tcp_rtx_queue) {
1525 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1526 if (err)
1527 return err;
1528 copied += skb->len;
1529 }
1530
1531 skb_queue_walk(&sk->sk_write_queue, skb) {
1532 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1533 if (err)
1534 break;
1535
1536 copied += skb->len;
1537 }
1538
1539 return err ?: copied;
1540}
1541
1542/* Clean up the receive buffer for full frames taken by the user,
1543 * then send an ACK if necessary. COPIED is the number of bytes
1544 * tcp_recvmsg has given to the user so far, it speeds up the
1545 * calculation of whether or not we must ACK for the sake of
1546 * a window update.
1547 */
1548void tcp_cleanup_rbuf(struct sock *sk, int copied)
1549{
1550 struct tcp_sock *tp = tcp_sk(sk);
1551 bool time_to_ack = false;
1552
1553 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1554
1555 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1556 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1557 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1558
1559 if (inet_csk_ack_scheduled(sk)) {
1560 const struct inet_connection_sock *icsk = inet_csk(sk);
1561
1562 if (/* Once-per-two-segments ACK was not sent by tcp_input.c */
1563 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1564 /*
1565 * If this read emptied read buffer, we send ACK, if
1566 * connection is not bidirectional, user drained
1567 * receive buffer and there was a small segment
1568 * in queue.
1569 */
1570 (copied > 0 &&
1571 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1572 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1573 !inet_csk_in_pingpong_mode(sk))) &&
1574 !atomic_read(&sk->sk_rmem_alloc)))
1575 time_to_ack = true;
1576 }
1577
1578 /* We send an ACK if we can now advertise a non-zero window
1579 * which has been raised "significantly".
1580 *
1581 * Even if window raised up to infinity, do not send window open ACK
1582 * in states, where we will not receive more. It is useless.
1583 */
1584 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1585 __u32 rcv_window_now = tcp_receive_window(tp);
1586
1587 /* Optimize, __tcp_select_window() is not cheap. */
1588 if (2*rcv_window_now <= tp->window_clamp) {
1589 __u32 new_window = __tcp_select_window(sk);
1590
1591 /* Send ACK now, if this read freed lots of space
1592 * in our buffer. Certainly, new_window is new window.
1593 * We can advertise it now, if it is not less than current one.
1594 * "Lots" means "at least twice" here.
1595 */
1596 if (new_window && new_window >= 2 * rcv_window_now)
1597 time_to_ack = true;
1598 }
1599 }
1600 if (time_to_ack)
1601 tcp_send_ack(sk);
1602}
1603
1604static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1605{
1606 struct sk_buff *skb;
1607 u32 offset;
1608
1609 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1610 offset = seq - TCP_SKB_CB(skb)->seq;
1611 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1612 pr_err_once("%s: found a SYN, please report !\n", __func__);
1613 offset--;
1614 }
1615 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1616 *off = offset;
1617 return skb;
1618 }
1619 /* This looks weird, but this can happen if TCP collapsing
1620 * splitted a fat GRO packet, while we released socket lock
1621 * in skb_splice_bits()
1622 */
1623 sk_eat_skb(sk, skb);
1624 }
1625 return NULL;
1626}
1627
1628/*
1629 * This routine provides an alternative to tcp_recvmsg() for routines
1630 * that would like to handle copying from skbuffs directly in 'sendfile'
1631 * fashion.
1632 * Note:
1633 * - It is assumed that the socket was locked by the caller.
1634 * - The routine does not block.
1635 * - At present, there is no support for reading OOB data
1636 * or for 'peeking' the socket using this routine
1637 * (although both would be easy to implement).
1638 */
1639int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1640 sk_read_actor_t recv_actor)
1641{
1642 struct sk_buff *skb;
1643 struct tcp_sock *tp = tcp_sk(sk);
1644 u32 seq = tp->copied_seq;
1645 u32 offset;
1646 int copied = 0;
1647
1648 if (sk->sk_state == TCP_LISTEN)
1649 return -ENOTCONN;
1650 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1651 if (offset < skb->len) {
1652 int used;
1653 size_t len;
1654
1655 len = skb->len - offset;
1656 /* Stop reading if we hit a patch of urgent data */
1657 if (tp->urg_data) {
1658 u32 urg_offset = tp->urg_seq - seq;
1659 if (urg_offset < len)
1660 len = urg_offset;
1661 if (!len)
1662 break;
1663 }
1664 used = recv_actor(desc, skb, offset, len);
1665 if (used <= 0) {
1666 if (!copied)
1667 copied = used;
1668 break;
1669 } else if (used <= len) {
1670 seq += used;
1671 copied += used;
1672 offset += used;
1673 }
1674 /* If recv_actor drops the lock (e.g. TCP splice
1675 * receive) the skb pointer might be invalid when
1676 * getting here: tcp_collapse might have deleted it
1677 * while aggregating skbs from the socket queue.
1678 */
1679 skb = tcp_recv_skb(sk, seq - 1, &offset);
1680 if (!skb)
1681 break;
1682 /* TCP coalescing might have appended data to the skb.
1683 * Try to splice more frags
1684 */
1685 if (offset + 1 != skb->len)
1686 continue;
1687 }
1688 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1689 sk_eat_skb(sk, skb);
1690 ++seq;
1691 break;
1692 }
1693 sk_eat_skb(sk, skb);
1694 if (!desc->count)
1695 break;
1696 WRITE_ONCE(tp->copied_seq, seq);
1697 }
1698 WRITE_ONCE(tp->copied_seq, seq);
1699
1700 tcp_rcv_space_adjust(sk);
1701
1702 /* Clean up data we have read: This will do ACK frames. */
1703 if (copied > 0) {
1704 tcp_recv_skb(sk, seq, &offset);
1705 tcp_cleanup_rbuf(sk, copied);
1706 }
1707 return copied;
1708}
1709EXPORT_SYMBOL(tcp_read_sock);
1710
1711int tcp_peek_len(struct socket *sock)
1712{
1713 return tcp_inq(sock->sk);
1714}
1715EXPORT_SYMBOL(tcp_peek_len);
1716
1717/* Make sure sk_rcvbuf is big enough to satisfy SO_RCVLOWAT hint */
1718int tcp_set_rcvlowat(struct sock *sk, int val)
1719{
1720 int cap;
1721
1722 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1723 cap = sk->sk_rcvbuf >> 1;
1724 else
1725 cap = sock_net(sk)->ipv4.sysctl_tcp_rmem[2] >> 1;
1726 val = min(val, cap);
1727 WRITE_ONCE(sk->sk_rcvlowat, val ? : 1);
1728
1729 /* Check if we need to signal EPOLLIN right now */
1730 tcp_data_ready(sk);
1731
1732 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK)
1733 return 0;
1734
1735 val <<= 1;
1736 if (val > sk->sk_rcvbuf) {
1737 WRITE_ONCE(sk->sk_rcvbuf, val);
1738 tcp_sk(sk)->window_clamp = tcp_win_from_space(sk, val);
1739 }
1740 return 0;
1741}
1742EXPORT_SYMBOL(tcp_set_rcvlowat);
1743
1744void tcp_update_recv_tstamps(struct sk_buff *skb,
1745 struct scm_timestamping_internal *tss)
1746{
1747 if (skb->tstamp)
1748 tss->ts[0] = ktime_to_timespec64(skb->tstamp);
1749 else
1750 tss->ts[0] = (struct timespec64) {0};
1751
1752 if (skb_hwtstamps(skb)->hwtstamp)
1753 tss->ts[2] = ktime_to_timespec64(skb_hwtstamps(skb)->hwtstamp);
1754 else
1755 tss->ts[2] = (struct timespec64) {0};
1756}
1757
1758#ifdef CONFIG_MMU
1759static const struct vm_operations_struct tcp_vm_ops = {
1760};
1761
1762int tcp_mmap(struct file *file, struct socket *sock,
1763 struct vm_area_struct *vma)
1764{
1765 if (vma->vm_flags & (VM_WRITE | VM_EXEC))
1766 return -EPERM;
1767 vma->vm_flags &= ~(VM_MAYWRITE | VM_MAYEXEC);
1768
1769 /* Instruct vm_insert_page() to not mmap_read_lock(mm) */
1770 vma->vm_flags |= VM_MIXEDMAP;
1771
1772 vma->vm_ops = &tcp_vm_ops;
1773 return 0;
1774}
1775EXPORT_SYMBOL(tcp_mmap);
1776
1777static skb_frag_t *skb_advance_to_frag(struct sk_buff *skb, u32 offset_skb,
1778 u32 *offset_frag)
1779{
1780 skb_frag_t *frag;
1781
1782 offset_skb -= skb_headlen(skb);
1783 if ((int)offset_skb < 0 || skb_has_frag_list(skb))
1784 return NULL;
1785
1786 frag = skb_shinfo(skb)->frags;
1787 while (offset_skb) {
1788 if (skb_frag_size(frag) > offset_skb) {
1789 *offset_frag = offset_skb;
1790 return frag;
1791 }
1792 offset_skb -= skb_frag_size(frag);
1793 ++frag;
1794 }
1795 *offset_frag = 0;
1796 return frag;
1797}
1798
1799static bool can_map_frag(const skb_frag_t *frag)
1800{
1801 return skb_frag_size(frag) == PAGE_SIZE && !skb_frag_off(frag);
1802}
1803
1804static int find_next_mappable_frag(const skb_frag_t *frag,
1805 int remaining_in_skb)
1806{
1807 int offset = 0;
1808
1809 if (likely(can_map_frag(frag)))
1810 return 0;
1811
1812 while (offset < remaining_in_skb && !can_map_frag(frag)) {
1813 offset += skb_frag_size(frag);
1814 ++frag;
1815 }
1816 return offset;
1817}
1818
1819static void tcp_zerocopy_set_hint_for_skb(struct sock *sk,
1820 struct tcp_zerocopy_receive *zc,
1821 struct sk_buff *skb, u32 offset)
1822{
1823 u32 frag_offset, partial_frag_remainder = 0;
1824 int mappable_offset;
1825 skb_frag_t *frag;
1826
1827 /* worst case: skip to next skb. try to improve on this case below */
1828 zc->recv_skip_hint = skb->len - offset;
1829
1830 /* Find the frag containing this offset (and how far into that frag) */
1831 frag = skb_advance_to_frag(skb, offset, &frag_offset);
1832 if (!frag)
1833 return;
1834
1835 if (frag_offset) {
1836 struct skb_shared_info *info = skb_shinfo(skb);
1837
1838 /* We read part of the last frag, must recvmsg() rest of skb. */
1839 if (frag == &info->frags[info->nr_frags - 1])
1840 return;
1841
1842 /* Else, we must at least read the remainder in this frag. */
1843 partial_frag_remainder = skb_frag_size(frag) - frag_offset;
1844 zc->recv_skip_hint -= partial_frag_remainder;
1845 ++frag;
1846 }
1847
1848 /* partial_frag_remainder: If part way through a frag, must read rest.
1849 * mappable_offset: Bytes till next mappable frag, *not* counting bytes
1850 * in partial_frag_remainder.
1851 */
1852 mappable_offset = find_next_mappable_frag(frag, zc->recv_skip_hint);
1853 zc->recv_skip_hint = mappable_offset + partial_frag_remainder;
1854}
1855
1856static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
1857 int nonblock, int flags,
1858 struct scm_timestamping_internal *tss,
1859 int *cmsg_flags);
1860static int receive_fallback_to_copy(struct sock *sk,
1861 struct tcp_zerocopy_receive *zc, int inq,
1862 struct scm_timestamping_internal *tss)
1863{
1864 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1865 struct msghdr msg = {};
1866 struct iovec iov;
1867 int err;
1868
1869 zc->length = 0;
1870 zc->recv_skip_hint = 0;
1871
1872 if (copy_address != zc->copybuf_address)
1873 return -EINVAL;
1874
1875 err = import_single_range(READ, (void __user *)copy_address,
1876 inq, &iov, &msg.msg_iter);
1877 if (err)
1878 return err;
1879
1880 err = tcp_recvmsg_locked(sk, &msg, inq, /*nonblock=*/1, /*flags=*/0,
1881 tss, &zc->msg_flags);
1882 if (err < 0)
1883 return err;
1884
1885 zc->copybuf_len = err;
1886 if (likely(zc->copybuf_len)) {
1887 struct sk_buff *skb;
1888 u32 offset;
1889
1890 skb = tcp_recv_skb(sk, tcp_sk(sk)->copied_seq, &offset);
1891 if (skb)
1892 tcp_zerocopy_set_hint_for_skb(sk, zc, skb, offset);
1893 }
1894 return 0;
1895}
1896
1897static int tcp_copy_straggler_data(struct tcp_zerocopy_receive *zc,
1898 struct sk_buff *skb, u32 copylen,
1899 u32 *offset, u32 *seq)
1900{
1901 unsigned long copy_address = (unsigned long)zc->copybuf_address;
1902 struct msghdr msg = {};
1903 struct iovec iov;
1904 int err;
1905
1906 if (copy_address != zc->copybuf_address)
1907 return -EINVAL;
1908
1909 err = import_single_range(READ, (void __user *)copy_address,
1910 copylen, &iov, &msg.msg_iter);
1911 if (err)
1912 return err;
1913 err = skb_copy_datagram_msg(skb, *offset, &msg, copylen);
1914 if (err)
1915 return err;
1916 zc->recv_skip_hint -= copylen;
1917 *offset += copylen;
1918 *seq += copylen;
1919 return (__s32)copylen;
1920}
1921
1922static int tcp_zc_handle_leftover(struct tcp_zerocopy_receive *zc,
1923 struct sock *sk,
1924 struct sk_buff *skb,
1925 u32 *seq,
1926 s32 copybuf_len,
1927 struct scm_timestamping_internal *tss)
1928{
1929 u32 offset, copylen = min_t(u32, copybuf_len, zc->recv_skip_hint);
1930
1931 if (!copylen)
1932 return 0;
1933 /* skb is null if inq < PAGE_SIZE. */
1934 if (skb) {
1935 offset = *seq - TCP_SKB_CB(skb)->seq;
1936 } else {
1937 skb = tcp_recv_skb(sk, *seq, &offset);
1938 if (TCP_SKB_CB(skb)->has_rxtstamp) {
1939 tcp_update_recv_tstamps(skb, tss);
1940 zc->msg_flags |= TCP_CMSG_TS;
1941 }
1942 }
1943
1944 zc->copybuf_len = tcp_copy_straggler_data(zc, skb, copylen, &offset,
1945 seq);
1946 return zc->copybuf_len < 0 ? 0 : copylen;
1947}
1948
1949static int tcp_zerocopy_vm_insert_batch_error(struct vm_area_struct *vma,
1950 struct page **pending_pages,
1951 unsigned long pages_remaining,
1952 unsigned long *address,
1953 u32 *length,
1954 u32 *seq,
1955 struct tcp_zerocopy_receive *zc,
1956 u32 total_bytes_to_map,
1957 int err)
1958{
1959 /* At least one page did not map. Try zapping if we skipped earlier. */
1960 if (err == -EBUSY &&
1961 zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT) {
1962 u32 maybe_zap_len;
1963
1964 maybe_zap_len = total_bytes_to_map - /* All bytes to map */
1965 *length + /* Mapped or pending */
1966 (pages_remaining * PAGE_SIZE); /* Failed map. */
1967 zap_page_range(vma, *address, maybe_zap_len);
1968 err = 0;
1969 }
1970
1971 if (!err) {
1972 unsigned long leftover_pages = pages_remaining;
1973 int bytes_mapped;
1974
1975 /* We called zap_page_range, try to reinsert. */
1976 err = vm_insert_pages(vma, *address,
1977 pending_pages,
1978 &pages_remaining);
1979 bytes_mapped = PAGE_SIZE * (leftover_pages - pages_remaining);
1980 *seq += bytes_mapped;
1981 *address += bytes_mapped;
1982 }
1983 if (err) {
1984 /* Either we were unable to zap, OR we zapped, retried an
1985 * insert, and still had an issue. Either ways, pages_remaining
1986 * is the number of pages we were unable to map, and we unroll
1987 * some state we speculatively touched before.
1988 */
1989 const int bytes_not_mapped = PAGE_SIZE * pages_remaining;
1990
1991 *length -= bytes_not_mapped;
1992 zc->recv_skip_hint += bytes_not_mapped;
1993 }
1994 return err;
1995}
1996
1997static int tcp_zerocopy_vm_insert_batch(struct vm_area_struct *vma,
1998 struct page **pages,
1999 unsigned int pages_to_map,
2000 unsigned long *address,
2001 u32 *length,
2002 u32 *seq,
2003 struct tcp_zerocopy_receive *zc,
2004 u32 total_bytes_to_map)
2005{
2006 unsigned long pages_remaining = pages_to_map;
2007 unsigned int pages_mapped;
2008 unsigned int bytes_mapped;
2009 int err;
2010
2011 err = vm_insert_pages(vma, *address, pages, &pages_remaining);
2012 pages_mapped = pages_to_map - (unsigned int)pages_remaining;
2013 bytes_mapped = PAGE_SIZE * pages_mapped;
2014 /* Even if vm_insert_pages fails, it may have partially succeeded in
2015 * mapping (some but not all of the pages).
2016 */
2017 *seq += bytes_mapped;
2018 *address += bytes_mapped;
2019
2020 if (likely(!err))
2021 return 0;
2022
2023 /* Error: maybe zap and retry + rollback state for failed inserts. */
2024 return tcp_zerocopy_vm_insert_batch_error(vma, pages + pages_mapped,
2025 pages_remaining, address, length, seq, zc, total_bytes_to_map,
2026 err);
2027}
2028
2029#define TCP_VALID_ZC_MSG_FLAGS (TCP_CMSG_TS)
2030static void tcp_zc_finalize_rx_tstamp(struct sock *sk,
2031 struct tcp_zerocopy_receive *zc,
2032 struct scm_timestamping_internal *tss)
2033{
2034 unsigned long msg_control_addr;
2035 struct msghdr cmsg_dummy;
2036
2037 msg_control_addr = (unsigned long)zc->msg_control;
2038 cmsg_dummy.msg_control = (void *)msg_control_addr;
2039 cmsg_dummy.msg_controllen =
2040 (__kernel_size_t)zc->msg_controllen;
2041 cmsg_dummy.msg_flags = in_compat_syscall()
2042 ? MSG_CMSG_COMPAT : 0;
2043 cmsg_dummy.msg_control_is_user = true;
2044 zc->msg_flags = 0;
2045 if (zc->msg_control == msg_control_addr &&
2046 zc->msg_controllen == cmsg_dummy.msg_controllen) {
2047 tcp_recv_timestamp(&cmsg_dummy, sk, tss);
2048 zc->msg_control = (__u64)
2049 ((uintptr_t)cmsg_dummy.msg_control);
2050 zc->msg_controllen =
2051 (__u64)cmsg_dummy.msg_controllen;
2052 zc->msg_flags = (__u32)cmsg_dummy.msg_flags;
2053 }
2054}
2055
2056#define TCP_ZEROCOPY_PAGE_BATCH_SIZE 32
2057static int tcp_zerocopy_receive(struct sock *sk,
2058 struct tcp_zerocopy_receive *zc,
2059 struct scm_timestamping_internal *tss)
2060{
2061 u32 length = 0, offset, vma_len, avail_len, copylen = 0;
2062 unsigned long address = (unsigned long)zc->address;
2063 struct page *pages[TCP_ZEROCOPY_PAGE_BATCH_SIZE];
2064 s32 copybuf_len = zc->copybuf_len;
2065 struct tcp_sock *tp = tcp_sk(sk);
2066 const skb_frag_t *frags = NULL;
2067 unsigned int pages_to_map = 0;
2068 struct vm_area_struct *vma;
2069 struct sk_buff *skb = NULL;
2070 u32 seq = tp->copied_seq;
2071 u32 total_bytes_to_map;
2072 int inq = tcp_inq(sk);
2073 int ret;
2074
2075 zc->copybuf_len = 0;
2076 zc->msg_flags = 0;
2077
2078 if (address & (PAGE_SIZE - 1) || address != zc->address)
2079 return -EINVAL;
2080
2081 if (sk->sk_state == TCP_LISTEN)
2082 return -ENOTCONN;
2083
2084 sock_rps_record_flow(sk);
2085
2086 if (inq && inq <= copybuf_len)
2087 return receive_fallback_to_copy(sk, zc, inq, tss);
2088
2089 if (inq < PAGE_SIZE) {
2090 zc->length = 0;
2091 zc->recv_skip_hint = inq;
2092 if (!inq && sock_flag(sk, SOCK_DONE))
2093 return -EIO;
2094 return 0;
2095 }
2096
2097 mmap_read_lock(current->mm);
2098
2099 vma = vma_lookup(current->mm, address);
2100 if (!vma || vma->vm_ops != &tcp_vm_ops) {
2101 mmap_read_unlock(current->mm);
2102 return -EINVAL;
2103 }
2104 vma_len = min_t(unsigned long, zc->length, vma->vm_end - address);
2105 avail_len = min_t(u32, vma_len, inq);
2106 total_bytes_to_map = avail_len & ~(PAGE_SIZE - 1);
2107 if (total_bytes_to_map) {
2108 if (!(zc->flags & TCP_RECEIVE_ZEROCOPY_FLAG_TLB_CLEAN_HINT))
2109 zap_page_range(vma, address, total_bytes_to_map);
2110 zc->length = total_bytes_to_map;
2111 zc->recv_skip_hint = 0;
2112 } else {
2113 zc->length = avail_len;
2114 zc->recv_skip_hint = avail_len;
2115 }
2116 ret = 0;
2117 while (length + PAGE_SIZE <= zc->length) {
2118 int mappable_offset;
2119 struct page *page;
2120
2121 if (zc->recv_skip_hint < PAGE_SIZE) {
2122 u32 offset_frag;
2123
2124 if (skb) {
2125 if (zc->recv_skip_hint > 0)
2126 break;
2127 skb = skb->next;
2128 offset = seq - TCP_SKB_CB(skb)->seq;
2129 } else {
2130 skb = tcp_recv_skb(sk, seq, &offset);
2131 }
2132
2133 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2134 tcp_update_recv_tstamps(skb, tss);
2135 zc->msg_flags |= TCP_CMSG_TS;
2136 }
2137 zc->recv_skip_hint = skb->len - offset;
2138 frags = skb_advance_to_frag(skb, offset, &offset_frag);
2139 if (!frags || offset_frag)
2140 break;
2141 }
2142
2143 mappable_offset = find_next_mappable_frag(frags,
2144 zc->recv_skip_hint);
2145 if (mappable_offset) {
2146 zc->recv_skip_hint = mappable_offset;
2147 break;
2148 }
2149 page = skb_frag_page(frags);
2150 prefetchw(page);
2151 pages[pages_to_map++] = page;
2152 length += PAGE_SIZE;
2153 zc->recv_skip_hint -= PAGE_SIZE;
2154 frags++;
2155 if (pages_to_map == TCP_ZEROCOPY_PAGE_BATCH_SIZE ||
2156 zc->recv_skip_hint < PAGE_SIZE) {
2157 /* Either full batch, or we're about to go to next skb
2158 * (and we cannot unroll failed ops across skbs).
2159 */
2160 ret = tcp_zerocopy_vm_insert_batch(vma, pages,
2161 pages_to_map,
2162 &address, &length,
2163 &seq, zc,
2164 total_bytes_to_map);
2165 if (ret)
2166 goto out;
2167 pages_to_map = 0;
2168 }
2169 }
2170 if (pages_to_map) {
2171 ret = tcp_zerocopy_vm_insert_batch(vma, pages, pages_to_map,
2172 &address, &length, &seq,
2173 zc, total_bytes_to_map);
2174 }
2175out:
2176 mmap_read_unlock(current->mm);
2177 /* Try to copy straggler data. */
2178 if (!ret)
2179 copylen = tcp_zc_handle_leftover(zc, sk, skb, &seq, copybuf_len, tss);
2180
2181 if (length + copylen) {
2182 WRITE_ONCE(tp->copied_seq, seq);
2183 tcp_rcv_space_adjust(sk);
2184
2185 /* Clean up data we have read: This will do ACK frames. */
2186 tcp_recv_skb(sk, seq, &offset);
2187 tcp_cleanup_rbuf(sk, length + copylen);
2188 ret = 0;
2189 if (length == zc->length)
2190 zc->recv_skip_hint = 0;
2191 } else {
2192 if (!zc->recv_skip_hint && sock_flag(sk, SOCK_DONE))
2193 ret = -EIO;
2194 }
2195 zc->length = length;
2196 return ret;
2197}
2198#endif
2199
2200/* Similar to __sock_recv_timestamp, but does not require an skb */
2201void tcp_recv_timestamp(struct msghdr *msg, const struct sock *sk,
2202 struct scm_timestamping_internal *tss)
2203{
2204 int new_tstamp = sock_flag(sk, SOCK_TSTAMP_NEW);
2205 bool has_timestamping = false;
2206
2207 if (tss->ts[0].tv_sec || tss->ts[0].tv_nsec) {
2208 if (sock_flag(sk, SOCK_RCVTSTAMP)) {
2209 if (sock_flag(sk, SOCK_RCVTSTAMPNS)) {
2210 if (new_tstamp) {
2211 struct __kernel_timespec kts = {
2212 .tv_sec = tss->ts[0].tv_sec,
2213 .tv_nsec = tss->ts[0].tv_nsec,
2214 };
2215 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_NEW,
2216 sizeof(kts), &kts);
2217 } else {
2218 struct __kernel_old_timespec ts_old = {
2219 .tv_sec = tss->ts[0].tv_sec,
2220 .tv_nsec = tss->ts[0].tv_nsec,
2221 };
2222 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMPNS_OLD,
2223 sizeof(ts_old), &ts_old);
2224 }
2225 } else {
2226 if (new_tstamp) {
2227 struct __kernel_sock_timeval stv = {
2228 .tv_sec = tss->ts[0].tv_sec,
2229 .tv_usec = tss->ts[0].tv_nsec / 1000,
2230 };
2231 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_NEW,
2232 sizeof(stv), &stv);
2233 } else {
2234 struct __kernel_old_timeval tv = {
2235 .tv_sec = tss->ts[0].tv_sec,
2236 .tv_usec = tss->ts[0].tv_nsec / 1000,
2237 };
2238 put_cmsg(msg, SOL_SOCKET, SO_TIMESTAMP_OLD,
2239 sizeof(tv), &tv);
2240 }
2241 }
2242 }
2243
2244 if (sk->sk_tsflags & SOF_TIMESTAMPING_SOFTWARE)
2245 has_timestamping = true;
2246 else
2247 tss->ts[0] = (struct timespec64) {0};
2248 }
2249
2250 if (tss->ts[2].tv_sec || tss->ts[2].tv_nsec) {
2251 if (sk->sk_tsflags & SOF_TIMESTAMPING_RAW_HARDWARE)
2252 has_timestamping = true;
2253 else
2254 tss->ts[2] = (struct timespec64) {0};
2255 }
2256
2257 if (has_timestamping) {
2258 tss->ts[1] = (struct timespec64) {0};
2259 if (sock_flag(sk, SOCK_TSTAMP_NEW))
2260 put_cmsg_scm_timestamping64(msg, tss);
2261 else
2262 put_cmsg_scm_timestamping(msg, tss);
2263 }
2264}
2265
2266static int tcp_inq_hint(struct sock *sk)
2267{
2268 const struct tcp_sock *tp = tcp_sk(sk);
2269 u32 copied_seq = READ_ONCE(tp->copied_seq);
2270 u32 rcv_nxt = READ_ONCE(tp->rcv_nxt);
2271 int inq;
2272
2273 inq = rcv_nxt - copied_seq;
2274 if (unlikely(inq < 0 || copied_seq != READ_ONCE(tp->copied_seq))) {
2275 lock_sock(sk);
2276 inq = tp->rcv_nxt - tp->copied_seq;
2277 release_sock(sk);
2278 }
2279 /* After receiving a FIN, tell the user-space to continue reading
2280 * by returning a non-zero inq.
2281 */
2282 if (inq == 0 && sock_flag(sk, SOCK_DONE))
2283 inq = 1;
2284 return inq;
2285}
2286
2287/*
2288 * This routine copies from a sock struct into the user buffer.
2289 *
2290 * Technical note: in 2.3 we work on _locked_ socket, so that
2291 * tricks with *seq access order and skb->users are not required.
2292 * Probably, code can be easily improved even more.
2293 */
2294
2295static int tcp_recvmsg_locked(struct sock *sk, struct msghdr *msg, size_t len,
2296 int nonblock, int flags,
2297 struct scm_timestamping_internal *tss,
2298 int *cmsg_flags)
2299{
2300 struct tcp_sock *tp = tcp_sk(sk);
2301 int copied = 0;
2302 u32 peek_seq;
2303 u32 *seq;
2304 unsigned long used;
2305 int err;
2306 int target; /* Read at least this many bytes */
2307 long timeo;
2308 struct sk_buff *skb, *last;
2309 u32 urg_hole = 0;
2310
2311 err = -ENOTCONN;
2312 if (sk->sk_state == TCP_LISTEN)
2313 goto out;
2314
2315 if (tp->recvmsg_inq)
2316 *cmsg_flags = TCP_CMSG_INQ;
2317 timeo = sock_rcvtimeo(sk, nonblock);
2318
2319 /* Urgent data needs to be handled specially. */
2320 if (flags & MSG_OOB)
2321 goto recv_urg;
2322
2323 if (unlikely(tp->repair)) {
2324 err = -EPERM;
2325 if (!(flags & MSG_PEEK))
2326 goto out;
2327
2328 if (tp->repair_queue == TCP_SEND_QUEUE)
2329 goto recv_sndq;
2330
2331 err = -EINVAL;
2332 if (tp->repair_queue == TCP_NO_QUEUE)
2333 goto out;
2334
2335 /* 'common' recv queue MSG_PEEK-ing */
2336 }
2337
2338 seq = &tp->copied_seq;
2339 if (flags & MSG_PEEK) {
2340 peek_seq = tp->copied_seq;
2341 seq = &peek_seq;
2342 }
2343
2344 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
2345
2346 do {
2347 u32 offset;
2348
2349 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
2350 if (tp->urg_data && tp->urg_seq == *seq) {
2351 if (copied)
2352 break;
2353 if (signal_pending(current)) {
2354 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
2355 break;
2356 }
2357 }
2358
2359 /* Next get a buffer. */
2360
2361 last = skb_peek_tail(&sk->sk_receive_queue);
2362 skb_queue_walk(&sk->sk_receive_queue, skb) {
2363 last = skb;
2364 /* Now that we have two receive queues this
2365 * shouldn't happen.
2366 */
2367 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
2368 "TCP recvmsg seq # bug: copied %X, seq %X, rcvnxt %X, fl %X\n",
2369 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
2370 flags))
2371 break;
2372
2373 offset = *seq - TCP_SKB_CB(skb)->seq;
2374 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2375 pr_err_once("%s: found a SYN, please report !\n", __func__);
2376 offset--;
2377 }
2378 if (offset < skb->len)
2379 goto found_ok_skb;
2380 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2381 goto found_fin_ok;
2382 WARN(!(flags & MSG_PEEK),
2383 "TCP recvmsg seq # bug 2: copied %X, seq %X, rcvnxt %X, fl %X\n",
2384 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
2385 }
2386
2387 /* Well, if we have backlog, try to process it now yet. */
2388
2389 if (copied >= target && !READ_ONCE(sk->sk_backlog.tail))
2390 break;
2391
2392 if (copied) {
2393 if (sk->sk_err ||
2394 sk->sk_state == TCP_CLOSE ||
2395 (sk->sk_shutdown & RCV_SHUTDOWN) ||
2396 !timeo ||
2397 signal_pending(current))
2398 break;
2399 } else {
2400 if (sock_flag(sk, SOCK_DONE))
2401 break;
2402
2403 if (sk->sk_err) {
2404 copied = sock_error(sk);
2405 break;
2406 }
2407
2408 if (sk->sk_shutdown & RCV_SHUTDOWN)
2409 break;
2410
2411 if (sk->sk_state == TCP_CLOSE) {
2412 /* This occurs when user tries to read
2413 * from never connected socket.
2414 */
2415 copied = -ENOTCONN;
2416 break;
2417 }
2418
2419 if (!timeo) {
2420 copied = -EAGAIN;
2421 break;
2422 }
2423
2424 if (signal_pending(current)) {
2425 copied = sock_intr_errno(timeo);
2426 break;
2427 }
2428 }
2429
2430 tcp_cleanup_rbuf(sk, copied);
2431
2432 if (copied >= target) {
2433 /* Do not sleep, just process backlog. */
2434 release_sock(sk);
2435 lock_sock(sk);
2436 } else {
2437 sk_wait_data(sk, &timeo, last);
2438 }
2439
2440 if ((flags & MSG_PEEK) &&
2441 (peek_seq - copied - urg_hole != tp->copied_seq)) {
2442 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
2443 current->comm,
2444 task_pid_nr(current));
2445 peek_seq = tp->copied_seq;
2446 }
2447 continue;
2448
2449found_ok_skb:
2450 /* Ok so how much can we use? */
2451 used = skb->len - offset;
2452 if (len < used)
2453 used = len;
2454
2455 /* Do we have urgent data here? */
2456 if (tp->urg_data) {
2457 u32 urg_offset = tp->urg_seq - *seq;
2458 if (urg_offset < used) {
2459 if (!urg_offset) {
2460 if (!sock_flag(sk, SOCK_URGINLINE)) {
2461 WRITE_ONCE(*seq, *seq + 1);
2462 urg_hole++;
2463 offset++;
2464 used--;
2465 if (!used)
2466 goto skip_copy;
2467 }
2468 } else
2469 used = urg_offset;
2470 }
2471 }
2472
2473 if (!(flags & MSG_TRUNC)) {
2474 err = skb_copy_datagram_msg(skb, offset, msg, used);
2475 if (err) {
2476 /* Exception. Bailout! */
2477 if (!copied)
2478 copied = -EFAULT;
2479 break;
2480 }
2481 }
2482
2483 WRITE_ONCE(*seq, *seq + used);
2484 copied += used;
2485 len -= used;
2486
2487 tcp_rcv_space_adjust(sk);
2488
2489skip_copy:
2490 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
2491 tp->urg_data = 0;
2492 tcp_fast_path_check(sk);
2493 }
2494
2495 if (TCP_SKB_CB(skb)->has_rxtstamp) {
2496 tcp_update_recv_tstamps(skb, tss);
2497 *cmsg_flags |= TCP_CMSG_TS;
2498 }
2499
2500 if (used + offset < skb->len)
2501 continue;
2502
2503 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2504 goto found_fin_ok;
2505 if (!(flags & MSG_PEEK))
2506 sk_eat_skb(sk, skb);
2507 continue;
2508
2509found_fin_ok:
2510 /* Process the FIN. */
2511 WRITE_ONCE(*seq, *seq + 1);
2512 if (!(flags & MSG_PEEK))
2513 sk_eat_skb(sk, skb);
2514 break;
2515 } while (len > 0);
2516
2517 /* According to UNIX98, msg_name/msg_namelen are ignored
2518 * on connected socket. I was just happy when found this 8) --ANK
2519 */
2520
2521 /* Clean up data we have read: This will do ACK frames. */
2522 tcp_cleanup_rbuf(sk, copied);
2523 return copied;
2524
2525out:
2526 return err;
2527
2528recv_urg:
2529 err = tcp_recv_urg(sk, msg, len, flags);
2530 goto out;
2531
2532recv_sndq:
2533 err = tcp_peek_sndq(sk, msg, len);
2534 goto out;
2535}
2536
2537int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
2538 int flags, int *addr_len)
2539{
2540 int cmsg_flags = 0, ret, inq;
2541 struct scm_timestamping_internal tss;
2542
2543 if (unlikely(flags & MSG_ERRQUEUE))
2544 return inet_recv_error(sk, msg, len, addr_len);
2545
2546 if (sk_can_busy_loop(sk) &&
2547 skb_queue_empty_lockless(&sk->sk_receive_queue) &&
2548 sk->sk_state == TCP_ESTABLISHED)
2549 sk_busy_loop(sk, nonblock);
2550
2551 lock_sock(sk);
2552 ret = tcp_recvmsg_locked(sk, msg, len, nonblock, flags, &tss,
2553 &cmsg_flags);
2554 release_sock(sk);
2555
2556 if (cmsg_flags && ret >= 0) {
2557 if (cmsg_flags & TCP_CMSG_TS)
2558 tcp_recv_timestamp(msg, sk, &tss);
2559 if (cmsg_flags & TCP_CMSG_INQ) {
2560 inq = tcp_inq_hint(sk);
2561 put_cmsg(msg, SOL_TCP, TCP_CM_INQ, sizeof(inq), &inq);
2562 }
2563 }
2564 return ret;
2565}
2566EXPORT_SYMBOL(tcp_recvmsg);
2567
2568void tcp_set_state(struct sock *sk, int state)
2569{
2570 int oldstate = sk->sk_state;
2571
2572 /* We defined a new enum for TCP states that are exported in BPF
2573 * so as not force the internal TCP states to be frozen. The
2574 * following checks will detect if an internal state value ever
2575 * differs from the BPF value. If this ever happens, then we will
2576 * need to remap the internal value to the BPF value before calling
2577 * tcp_call_bpf_2arg.
2578 */
2579 BUILD_BUG_ON((int)BPF_TCP_ESTABLISHED != (int)TCP_ESTABLISHED);
2580 BUILD_BUG_ON((int)BPF_TCP_SYN_SENT != (int)TCP_SYN_SENT);
2581 BUILD_BUG_ON((int)BPF_TCP_SYN_RECV != (int)TCP_SYN_RECV);
2582 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT1 != (int)TCP_FIN_WAIT1);
2583 BUILD_BUG_ON((int)BPF_TCP_FIN_WAIT2 != (int)TCP_FIN_WAIT2);
2584 BUILD_BUG_ON((int)BPF_TCP_TIME_WAIT != (int)TCP_TIME_WAIT);
2585 BUILD_BUG_ON((int)BPF_TCP_CLOSE != (int)TCP_CLOSE);
2586 BUILD_BUG_ON((int)BPF_TCP_CLOSE_WAIT != (int)TCP_CLOSE_WAIT);
2587 BUILD_BUG_ON((int)BPF_TCP_LAST_ACK != (int)TCP_LAST_ACK);
2588 BUILD_BUG_ON((int)BPF_TCP_LISTEN != (int)TCP_LISTEN);
2589 BUILD_BUG_ON((int)BPF_TCP_CLOSING != (int)TCP_CLOSING);
2590 BUILD_BUG_ON((int)BPF_TCP_NEW_SYN_RECV != (int)TCP_NEW_SYN_RECV);
2591 BUILD_BUG_ON((int)BPF_TCP_MAX_STATES != (int)TCP_MAX_STATES);
2592
2593 /* bpf uapi header bpf.h defines an anonymous enum with values
2594 * BPF_TCP_* used by bpf programs. Currently gcc built vmlinux
2595 * is able to emit this enum in DWARF due to the above BUILD_BUG_ON.
2596 * But clang built vmlinux does not have this enum in DWARF
2597 * since clang removes the above code before generating IR/debuginfo.
2598 * Let us explicitly emit the type debuginfo to ensure the
2599 * above-mentioned anonymous enum in the vmlinux DWARF and hence BTF
2600 * regardless of which compiler is used.
2601 */
2602 BTF_TYPE_EMIT_ENUM(BPF_TCP_ESTABLISHED);
2603
2604 if (BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), BPF_SOCK_OPS_STATE_CB_FLAG))
2605 tcp_call_bpf_2arg(sk, BPF_SOCK_OPS_STATE_CB, oldstate, state);
2606
2607 switch (state) {
2608 case TCP_ESTABLISHED:
2609 if (oldstate != TCP_ESTABLISHED)
2610 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2611 break;
2612
2613 case TCP_CLOSE:
2614 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
2615 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
2616
2617 sk->sk_prot->unhash(sk);
2618 if (inet_csk(sk)->icsk_bind_hash &&
2619 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
2620 inet_put_port(sk);
2621 fallthrough;
2622 default:
2623 if (oldstate == TCP_ESTABLISHED)
2624 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
2625 }
2626
2627 /* Change state AFTER socket is unhashed to avoid closed
2628 * socket sitting in hash tables.
2629 */
2630 inet_sk_state_store(sk, state);
2631}
2632EXPORT_SYMBOL_GPL(tcp_set_state);
2633
2634/*
2635 * State processing on a close. This implements the state shift for
2636 * sending our FIN frame. Note that we only send a FIN for some
2637 * states. A shutdown() may have already sent the FIN, or we may be
2638 * closed.
2639 */
2640
2641static const unsigned char new_state[16] = {
2642 /* current state: new state: action: */
2643 [0 /* (Invalid) */] = TCP_CLOSE,
2644 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2645 [TCP_SYN_SENT] = TCP_CLOSE,
2646 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2647 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2648 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2649 [TCP_TIME_WAIT] = TCP_CLOSE,
2650 [TCP_CLOSE] = TCP_CLOSE,
2651 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2652 [TCP_LAST_ACK] = TCP_LAST_ACK,
2653 [TCP_LISTEN] = TCP_CLOSE,
2654 [TCP_CLOSING] = TCP_CLOSING,
2655 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2656};
2657
2658static int tcp_close_state(struct sock *sk)
2659{
2660 int next = (int)new_state[sk->sk_state];
2661 int ns = next & TCP_STATE_MASK;
2662
2663 tcp_set_state(sk, ns);
2664
2665 return next & TCP_ACTION_FIN;
2666}
2667
2668/*
2669 * Shutdown the sending side of a connection. Much like close except
2670 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2671 */
2672
2673void tcp_shutdown(struct sock *sk, int how)
2674{
2675 /* We need to grab some memory, and put together a FIN,
2676 * and then put it into the queue to be sent.
2677 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2678 */
2679 if (!(how & SEND_SHUTDOWN))
2680 return;
2681
2682 /* If we've already sent a FIN, or it's a closed state, skip this. */
2683 if ((1 << sk->sk_state) &
2684 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2685 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2686 /* Clear out any half completed packets. FIN if needed. */
2687 if (tcp_close_state(sk))
2688 tcp_send_fin(sk);
2689 }
2690}
2691EXPORT_SYMBOL(tcp_shutdown);
2692
2693bool tcp_check_oom(struct sock *sk, int shift)
2694{
2695 bool too_many_orphans, out_of_socket_memory;
2696
2697 too_many_orphans = tcp_too_many_orphans(sk, shift);
2698 out_of_socket_memory = tcp_out_of_memory(sk);
2699
2700 if (too_many_orphans)
2701 net_info_ratelimited("too many orphaned sockets\n");
2702 if (out_of_socket_memory)
2703 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2704 return too_many_orphans || out_of_socket_memory;
2705}
2706
2707void __tcp_close(struct sock *sk, long timeout)
2708{
2709 struct sk_buff *skb;
2710 int data_was_unread = 0;
2711 int state;
2712
2713 sk->sk_shutdown = SHUTDOWN_MASK;
2714
2715 if (sk->sk_state == TCP_LISTEN) {
2716 tcp_set_state(sk, TCP_CLOSE);
2717
2718 /* Special case. */
2719 inet_csk_listen_stop(sk);
2720
2721 goto adjudge_to_death;
2722 }
2723
2724 /* We need to flush the recv. buffs. We do this only on the
2725 * descriptor close, not protocol-sourced closes, because the
2726 * reader process may not have drained the data yet!
2727 */
2728 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2729 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2730
2731 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2732 len--;
2733 data_was_unread += len;
2734 __kfree_skb(skb);
2735 }
2736
2737 sk_mem_reclaim(sk);
2738
2739 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2740 if (sk->sk_state == TCP_CLOSE)
2741 goto adjudge_to_death;
2742
2743 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2744 * data was lost. To witness the awful effects of the old behavior of
2745 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2746 * GET in an FTP client, suspend the process, wait for the client to
2747 * advertise a zero window, then kill -9 the FTP client, wheee...
2748 * Note: timeout is always zero in such a case.
2749 */
2750 if (unlikely(tcp_sk(sk)->repair)) {
2751 sk->sk_prot->disconnect(sk, 0);
2752 } else if (data_was_unread) {
2753 /* Unread data was tossed, zap the connection. */
2754 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2755 tcp_set_state(sk, TCP_CLOSE);
2756 tcp_send_active_reset(sk, sk->sk_allocation);
2757 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2758 /* Check zero linger _after_ checking for unread data. */
2759 sk->sk_prot->disconnect(sk, 0);
2760 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2761 } else if (tcp_close_state(sk)) {
2762 /* We FIN if the application ate all the data before
2763 * zapping the connection.
2764 */
2765
2766 /* RED-PEN. Formally speaking, we have broken TCP state
2767 * machine. State transitions:
2768 *
2769 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2770 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2771 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2772 *
2773 * are legal only when FIN has been sent (i.e. in window),
2774 * rather than queued out of window. Purists blame.
2775 *
2776 * F.e. "RFC state" is ESTABLISHED,
2777 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2778 *
2779 * The visible declinations are that sometimes
2780 * we enter time-wait state, when it is not required really
2781 * (harmless), do not send active resets, when they are
2782 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2783 * they look as CLOSING or LAST_ACK for Linux)
2784 * Probably, I missed some more holelets.
2785 * --ANK
2786 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2787 * in a single packet! (May consider it later but will
2788 * probably need API support or TCP_CORK SYN-ACK until
2789 * data is written and socket is closed.)
2790 */
2791 tcp_send_fin(sk);
2792 }
2793
2794 sk_stream_wait_close(sk, timeout);
2795
2796adjudge_to_death:
2797 state = sk->sk_state;
2798 sock_hold(sk);
2799 sock_orphan(sk);
2800
2801 local_bh_disable();
2802 bh_lock_sock(sk);
2803 /* remove backlog if any, without releasing ownership. */
2804 __release_sock(sk);
2805
2806 percpu_counter_inc(sk->sk_prot->orphan_count);
2807
2808 /* Have we already been destroyed by a softirq or backlog? */
2809 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2810 goto out;
2811
2812 /* This is a (useful) BSD violating of the RFC. There is a
2813 * problem with TCP as specified in that the other end could
2814 * keep a socket open forever with no application left this end.
2815 * We use a 1 minute timeout (about the same as BSD) then kill
2816 * our end. If they send after that then tough - BUT: long enough
2817 * that we won't make the old 4*rto = almost no time - whoops
2818 * reset mistake.
2819 *
2820 * Nope, it was not mistake. It is really desired behaviour
2821 * f.e. on http servers, when such sockets are useless, but
2822 * consume significant resources. Let's do it with special
2823 * linger2 option. --ANK
2824 */
2825
2826 if (sk->sk_state == TCP_FIN_WAIT2) {
2827 struct tcp_sock *tp = tcp_sk(sk);
2828 if (tp->linger2 < 0) {
2829 tcp_set_state(sk, TCP_CLOSE);
2830 tcp_send_active_reset(sk, GFP_ATOMIC);
2831 __NET_INC_STATS(sock_net(sk),
2832 LINUX_MIB_TCPABORTONLINGER);
2833 } else {
2834 const int tmo = tcp_fin_time(sk);
2835
2836 if (tmo > TCP_TIMEWAIT_LEN) {
2837 inet_csk_reset_keepalive_timer(sk,
2838 tmo - TCP_TIMEWAIT_LEN);
2839 } else {
2840 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2841 goto out;
2842 }
2843 }
2844 }
2845 if (sk->sk_state != TCP_CLOSE) {
2846 sk_mem_reclaim(sk);
2847 if (tcp_check_oom(sk, 0)) {
2848 tcp_set_state(sk, TCP_CLOSE);
2849 tcp_send_active_reset(sk, GFP_ATOMIC);
2850 __NET_INC_STATS(sock_net(sk),
2851 LINUX_MIB_TCPABORTONMEMORY);
2852 } else if (!check_net(sock_net(sk))) {
2853 /* Not possible to send reset; just close */
2854 tcp_set_state(sk, TCP_CLOSE);
2855 }
2856 }
2857
2858 if (sk->sk_state == TCP_CLOSE) {
2859 struct request_sock *req;
2860
2861 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk,
2862 lockdep_sock_is_held(sk));
2863 /* We could get here with a non-NULL req if the socket is
2864 * aborted (e.g., closed with unread data) before 3WHS
2865 * finishes.
2866 */
2867 if (req)
2868 reqsk_fastopen_remove(sk, req, false);
2869 inet_csk_destroy_sock(sk);
2870 }
2871 /* Otherwise, socket is reprieved until protocol close. */
2872
2873out:
2874 bh_unlock_sock(sk);
2875 local_bh_enable();
2876}
2877
2878void tcp_close(struct sock *sk, long timeout)
2879{
2880 lock_sock(sk);
2881 __tcp_close(sk, timeout);
2882 release_sock(sk);
2883 sock_put(sk);
2884}
2885EXPORT_SYMBOL(tcp_close);
2886
2887/* These states need RST on ABORT according to RFC793 */
2888
2889static inline bool tcp_need_reset(int state)
2890{
2891 return (1 << state) &
2892 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2893 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2894}
2895
2896static void tcp_rtx_queue_purge(struct sock *sk)
2897{
2898 struct rb_node *p = rb_first(&sk->tcp_rtx_queue);
2899
2900 tcp_sk(sk)->highest_sack = NULL;
2901 while (p) {
2902 struct sk_buff *skb = rb_to_skb(p);
2903
2904 p = rb_next(p);
2905 /* Since we are deleting whole queue, no need to
2906 * list_del(&skb->tcp_tsorted_anchor)
2907 */
2908 tcp_rtx_queue_unlink(skb, sk);
2909 sk_wmem_free_skb(sk, skb);
2910 }
2911}
2912
2913void tcp_write_queue_purge(struct sock *sk)
2914{
2915 struct sk_buff *skb;
2916
2917 tcp_chrono_stop(sk, TCP_CHRONO_BUSY);
2918 while ((skb = __skb_dequeue(&sk->sk_write_queue)) != NULL) {
2919 tcp_skb_tsorted_anchor_cleanup(skb);
2920 sk_wmem_free_skb(sk, skb);
2921 }
2922 tcp_rtx_queue_purge(sk);
2923 skb = sk->sk_tx_skb_cache;
2924 if (skb) {
2925 __kfree_skb(skb);
2926 sk->sk_tx_skb_cache = NULL;
2927 }
2928 INIT_LIST_HEAD(&tcp_sk(sk)->tsorted_sent_queue);
2929 sk_mem_reclaim(sk);
2930 tcp_clear_all_retrans_hints(tcp_sk(sk));
2931 tcp_sk(sk)->packets_out = 0;
2932 inet_csk(sk)->icsk_backoff = 0;
2933}
2934
2935int tcp_disconnect(struct sock *sk, int flags)
2936{
2937 struct inet_sock *inet = inet_sk(sk);
2938 struct inet_connection_sock *icsk = inet_csk(sk);
2939 struct tcp_sock *tp = tcp_sk(sk);
2940 int old_state = sk->sk_state;
2941 u32 seq;
2942
2943 if (old_state != TCP_CLOSE)
2944 tcp_set_state(sk, TCP_CLOSE);
2945
2946 /* ABORT function of RFC793 */
2947 if (old_state == TCP_LISTEN) {
2948 inet_csk_listen_stop(sk);
2949 } else if (unlikely(tp->repair)) {
2950 sk->sk_err = ECONNABORTED;
2951 } else if (tcp_need_reset(old_state) ||
2952 (tp->snd_nxt != tp->write_seq &&
2953 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2954 /* The last check adjusts for discrepancy of Linux wrt. RFC
2955 * states
2956 */
2957 tcp_send_active_reset(sk, gfp_any());
2958 sk->sk_err = ECONNRESET;
2959 } else if (old_state == TCP_SYN_SENT)
2960 sk->sk_err = ECONNRESET;
2961
2962 tcp_clear_xmit_timers(sk);
2963 __skb_queue_purge(&sk->sk_receive_queue);
2964 if (sk->sk_rx_skb_cache) {
2965 __kfree_skb(sk->sk_rx_skb_cache);
2966 sk->sk_rx_skb_cache = NULL;
2967 }
2968 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
2969 tp->urg_data = 0;
2970 tcp_write_queue_purge(sk);
2971 tcp_fastopen_active_disable_ofo_check(sk);
2972 skb_rbtree_purge(&tp->out_of_order_queue);
2973
2974 inet->inet_dport = 0;
2975
2976 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2977 inet_reset_saddr(sk);
2978
2979 sk->sk_shutdown = 0;
2980 sock_reset_flag(sk, SOCK_DONE);
2981 tp->srtt_us = 0;
2982 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
2983 tp->rcv_rtt_last_tsecr = 0;
2984
2985 seq = tp->write_seq + tp->max_window + 2;
2986 if (!seq)
2987 seq = 1;
2988 WRITE_ONCE(tp->write_seq, seq);
2989
2990 icsk->icsk_backoff = 0;
2991 icsk->icsk_probes_out = 0;
2992 icsk->icsk_probes_tstamp = 0;
2993 icsk->icsk_rto = TCP_TIMEOUT_INIT;
2994 icsk->icsk_rto_min = TCP_RTO_MIN;
2995 icsk->icsk_delack_max = TCP_DELACK_MAX;
2996 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2997 tp->snd_cwnd = TCP_INIT_CWND;
2998 tp->snd_cwnd_cnt = 0;
2999 tp->window_clamp = 0;
3000 tp->delivered = 0;
3001 tp->delivered_ce = 0;
3002 if (icsk->icsk_ca_ops->release)
3003 icsk->icsk_ca_ops->release(sk);
3004 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
3005 icsk->icsk_ca_initialized = 0;
3006 tcp_set_ca_state(sk, TCP_CA_Open);
3007 tp->is_sack_reneg = 0;
3008 tcp_clear_retrans(tp);
3009 tp->total_retrans = 0;
3010 inet_csk_delack_init(sk);
3011 /* Initialize rcv_mss to TCP_MIN_MSS to avoid division by 0
3012 * issue in __tcp_select_window()
3013 */
3014 icsk->icsk_ack.rcv_mss = TCP_MIN_MSS;
3015 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
3016 __sk_dst_reset(sk);
3017 dst_release(sk->sk_rx_dst);
3018 sk->sk_rx_dst = NULL;
3019 tcp_saved_syn_free(tp);
3020 tp->compressed_ack = 0;
3021 tp->segs_in = 0;
3022 tp->segs_out = 0;
3023 tp->bytes_sent = 0;
3024 tp->bytes_acked = 0;
3025 tp->bytes_received = 0;
3026 tp->bytes_retrans = 0;
3027 tp->data_segs_in = 0;
3028 tp->data_segs_out = 0;
3029 tp->duplicate_sack[0].start_seq = 0;
3030 tp->duplicate_sack[0].end_seq = 0;
3031 tp->dsack_dups = 0;
3032 tp->reord_seen = 0;
3033 tp->retrans_out = 0;
3034 tp->sacked_out = 0;
3035 tp->tlp_high_seq = 0;
3036 tp->last_oow_ack_time = 0;
3037 /* There's a bubble in the pipe until at least the first ACK. */
3038 tp->app_limited = ~0U;
3039 tp->rack.mstamp = 0;
3040 tp->rack.advanced = 0;
3041 tp->rack.reo_wnd_steps = 1;
3042 tp->rack.last_delivered = 0;
3043 tp->rack.reo_wnd_persist = 0;
3044 tp->rack.dsack_seen = 0;
3045 tp->syn_data_acked = 0;
3046 tp->rx_opt.saw_tstamp = 0;
3047 tp->rx_opt.dsack = 0;
3048 tp->rx_opt.num_sacks = 0;
3049 tp->rcv_ooopack = 0;
3050
3051
3052 /* Clean up fastopen related fields */
3053 tcp_free_fastopen_req(tp);
3054 inet->defer_connect = 0;
3055 tp->fastopen_client_fail = 0;
3056
3057 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
3058
3059 if (sk->sk_frag.page) {
3060 put_page(sk->sk_frag.page);
3061 sk->sk_frag.page = NULL;
3062 sk->sk_frag.offset = 0;
3063 }
3064
3065 sk_error_report(sk);
3066 return 0;
3067}
3068EXPORT_SYMBOL(tcp_disconnect);
3069
3070static inline bool tcp_can_repair_sock(const struct sock *sk)
3071{
3072 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
3073 (sk->sk_state != TCP_LISTEN);
3074}
3075
3076static int tcp_repair_set_window(struct tcp_sock *tp, sockptr_t optbuf, int len)
3077{
3078 struct tcp_repair_window opt;
3079
3080 if (!tp->repair)
3081 return -EPERM;
3082
3083 if (len != sizeof(opt))
3084 return -EINVAL;
3085
3086 if (copy_from_sockptr(&opt, optbuf, sizeof(opt)))
3087 return -EFAULT;
3088
3089 if (opt.max_window < opt.snd_wnd)
3090 return -EINVAL;
3091
3092 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
3093 return -EINVAL;
3094
3095 if (after(opt.rcv_wup, tp->rcv_nxt))
3096 return -EINVAL;
3097
3098 tp->snd_wl1 = opt.snd_wl1;
3099 tp->snd_wnd = opt.snd_wnd;
3100 tp->max_window = opt.max_window;
3101
3102 tp->rcv_wnd = opt.rcv_wnd;
3103 tp->rcv_wup = opt.rcv_wup;
3104
3105 return 0;
3106}
3107
3108static int tcp_repair_options_est(struct sock *sk, sockptr_t optbuf,
3109 unsigned int len)
3110{
3111 struct tcp_sock *tp = tcp_sk(sk);
3112 struct tcp_repair_opt opt;
3113 size_t offset = 0;
3114
3115 while (len >= sizeof(opt)) {
3116 if (copy_from_sockptr_offset(&opt, optbuf, offset, sizeof(opt)))
3117 return -EFAULT;
3118
3119 offset += sizeof(opt);
3120 len -= sizeof(opt);
3121
3122 switch (opt.opt_code) {
3123 case TCPOPT_MSS:
3124 tp->rx_opt.mss_clamp = opt.opt_val;
3125 tcp_mtup_init(sk);
3126 break;
3127 case TCPOPT_WINDOW:
3128 {
3129 u16 snd_wscale = opt.opt_val & 0xFFFF;
3130 u16 rcv_wscale = opt.opt_val >> 16;
3131
3132 if (snd_wscale > TCP_MAX_WSCALE || rcv_wscale > TCP_MAX_WSCALE)
3133 return -EFBIG;
3134
3135 tp->rx_opt.snd_wscale = snd_wscale;
3136 tp->rx_opt.rcv_wscale = rcv_wscale;
3137 tp->rx_opt.wscale_ok = 1;
3138 }
3139 break;
3140 case TCPOPT_SACK_PERM:
3141 if (opt.opt_val != 0)
3142 return -EINVAL;
3143
3144 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
3145 break;
3146 case TCPOPT_TIMESTAMP:
3147 if (opt.opt_val != 0)
3148 return -EINVAL;
3149
3150 tp->rx_opt.tstamp_ok = 1;
3151 break;
3152 }
3153 }
3154
3155 return 0;
3156}
3157
3158DEFINE_STATIC_KEY_FALSE(tcp_tx_delay_enabled);
3159EXPORT_SYMBOL(tcp_tx_delay_enabled);
3160
3161static void tcp_enable_tx_delay(void)
3162{
3163 if (!static_branch_unlikely(&tcp_tx_delay_enabled)) {
3164 static int __tcp_tx_delay_enabled = 0;
3165
3166 if (cmpxchg(&__tcp_tx_delay_enabled, 0, 1) == 0) {
3167 static_branch_enable(&tcp_tx_delay_enabled);
3168 pr_info("TCP_TX_DELAY enabled\n");
3169 }
3170 }
3171}
3172
3173/* When set indicates to always queue non-full frames. Later the user clears
3174 * this option and we transmit any pending partial frames in the queue. This is
3175 * meant to be used alongside sendfile() to get properly filled frames when the
3176 * user (for example) must write out headers with a write() call first and then
3177 * use sendfile to send out the data parts.
3178 *
3179 * TCP_CORK can be set together with TCP_NODELAY and it is stronger than
3180 * TCP_NODELAY.
3181 */
3182static void __tcp_sock_set_cork(struct sock *sk, bool on)
3183{
3184 struct tcp_sock *tp = tcp_sk(sk);
3185
3186 if (on) {
3187 tp->nonagle |= TCP_NAGLE_CORK;
3188 } else {
3189 tp->nonagle &= ~TCP_NAGLE_CORK;
3190 if (tp->nonagle & TCP_NAGLE_OFF)
3191 tp->nonagle |= TCP_NAGLE_PUSH;
3192 tcp_push_pending_frames(sk);
3193 }
3194}
3195
3196void tcp_sock_set_cork(struct sock *sk, bool on)
3197{
3198 lock_sock(sk);
3199 __tcp_sock_set_cork(sk, on);
3200 release_sock(sk);
3201}
3202EXPORT_SYMBOL(tcp_sock_set_cork);
3203
3204/* TCP_NODELAY is weaker than TCP_CORK, so that this option on corked socket is
3205 * remembered, but it is not activated until cork is cleared.
3206 *
3207 * However, when TCP_NODELAY is set we make an explicit push, which overrides
3208 * even TCP_CORK for currently queued segments.
3209 */
3210static void __tcp_sock_set_nodelay(struct sock *sk, bool on)
3211{
3212 if (on) {
3213 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
3214 tcp_push_pending_frames(sk);
3215 } else {
3216 tcp_sk(sk)->nonagle &= ~TCP_NAGLE_OFF;
3217 }
3218}
3219
3220void tcp_sock_set_nodelay(struct sock *sk)
3221{
3222 lock_sock(sk);
3223 __tcp_sock_set_nodelay(sk, true);
3224 release_sock(sk);
3225}
3226EXPORT_SYMBOL(tcp_sock_set_nodelay);
3227
3228static void __tcp_sock_set_quickack(struct sock *sk, int val)
3229{
3230 if (!val) {
3231 inet_csk_enter_pingpong_mode(sk);
3232 return;
3233 }
3234
3235 inet_csk_exit_pingpong_mode(sk);
3236 if ((1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
3237 inet_csk_ack_scheduled(sk)) {
3238 inet_csk(sk)->icsk_ack.pending |= ICSK_ACK_PUSHED;
3239 tcp_cleanup_rbuf(sk, 1);
3240 if (!(val & 1))
3241 inet_csk_enter_pingpong_mode(sk);
3242 }
3243}
3244
3245void tcp_sock_set_quickack(struct sock *sk, int val)
3246{
3247 lock_sock(sk);
3248 __tcp_sock_set_quickack(sk, val);
3249 release_sock(sk);
3250}
3251EXPORT_SYMBOL(tcp_sock_set_quickack);
3252
3253int tcp_sock_set_syncnt(struct sock *sk, int val)
3254{
3255 if (val < 1 || val > MAX_TCP_SYNCNT)
3256 return -EINVAL;
3257
3258 lock_sock(sk);
3259 inet_csk(sk)->icsk_syn_retries = val;
3260 release_sock(sk);
3261 return 0;
3262}
3263EXPORT_SYMBOL(tcp_sock_set_syncnt);
3264
3265void tcp_sock_set_user_timeout(struct sock *sk, u32 val)
3266{
3267 lock_sock(sk);
3268 inet_csk(sk)->icsk_user_timeout = val;
3269 release_sock(sk);
3270}
3271EXPORT_SYMBOL(tcp_sock_set_user_timeout);
3272
3273int tcp_sock_set_keepidle_locked(struct sock *sk, int val)
3274{
3275 struct tcp_sock *tp = tcp_sk(sk);
3276
3277 if (val < 1 || val > MAX_TCP_KEEPIDLE)
3278 return -EINVAL;
3279
3280 tp->keepalive_time = val * HZ;
3281 if (sock_flag(sk, SOCK_KEEPOPEN) &&
3282 !((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))) {
3283 u32 elapsed = keepalive_time_elapsed(tp);
3284
3285 if (tp->keepalive_time > elapsed)
3286 elapsed = tp->keepalive_time - elapsed;
3287 else
3288 elapsed = 0;
3289 inet_csk_reset_keepalive_timer(sk, elapsed);
3290 }
3291
3292 return 0;
3293}
3294
3295int tcp_sock_set_keepidle(struct sock *sk, int val)
3296{
3297 int err;
3298
3299 lock_sock(sk);
3300 err = tcp_sock_set_keepidle_locked(sk, val);
3301 release_sock(sk);
3302 return err;
3303}
3304EXPORT_SYMBOL(tcp_sock_set_keepidle);
3305
3306int tcp_sock_set_keepintvl(struct sock *sk, int val)
3307{
3308 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3309 return -EINVAL;
3310
3311 lock_sock(sk);
3312 tcp_sk(sk)->keepalive_intvl = val * HZ;
3313 release_sock(sk);
3314 return 0;
3315}
3316EXPORT_SYMBOL(tcp_sock_set_keepintvl);
3317
3318int tcp_sock_set_keepcnt(struct sock *sk, int val)
3319{
3320 if (val < 1 || val > MAX_TCP_KEEPCNT)
3321 return -EINVAL;
3322
3323 lock_sock(sk);
3324 tcp_sk(sk)->keepalive_probes = val;
3325 release_sock(sk);
3326 return 0;
3327}
3328EXPORT_SYMBOL(tcp_sock_set_keepcnt);
3329
3330int tcp_set_window_clamp(struct sock *sk, int val)
3331{
3332 struct tcp_sock *tp = tcp_sk(sk);
3333
3334 if (!val) {
3335 if (sk->sk_state != TCP_CLOSE)
3336 return -EINVAL;
3337 tp->window_clamp = 0;
3338 } else {
3339 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
3340 SOCK_MIN_RCVBUF / 2 : val;
3341 }
3342 return 0;
3343}
3344
3345/*
3346 * Socket option code for TCP.
3347 */
3348static int do_tcp_setsockopt(struct sock *sk, int level, int optname,
3349 sockptr_t optval, unsigned int optlen)
3350{
3351 struct tcp_sock *tp = tcp_sk(sk);
3352 struct inet_connection_sock *icsk = inet_csk(sk);
3353 struct net *net = sock_net(sk);
3354 int val;
3355 int err = 0;
3356
3357 /* These are data/string values, all the others are ints */
3358 switch (optname) {
3359 case TCP_CONGESTION: {
3360 char name[TCP_CA_NAME_MAX];
3361
3362 if (optlen < 1)
3363 return -EINVAL;
3364
3365 val = strncpy_from_sockptr(name, optval,
3366 min_t(long, TCP_CA_NAME_MAX-1, optlen));
3367 if (val < 0)
3368 return -EFAULT;
3369 name[val] = 0;
3370
3371 lock_sock(sk);
3372 err = tcp_set_congestion_control(sk, name, true,
3373 ns_capable(sock_net(sk)->user_ns,
3374 CAP_NET_ADMIN));
3375 release_sock(sk);
3376 return err;
3377 }
3378 case TCP_ULP: {
3379 char name[TCP_ULP_NAME_MAX];
3380
3381 if (optlen < 1)
3382 return -EINVAL;
3383
3384 val = strncpy_from_sockptr(name, optval,
3385 min_t(long, TCP_ULP_NAME_MAX - 1,
3386 optlen));
3387 if (val < 0)
3388 return -EFAULT;
3389 name[val] = 0;
3390
3391 lock_sock(sk);
3392 err = tcp_set_ulp(sk, name);
3393 release_sock(sk);
3394 return err;
3395 }
3396 case TCP_FASTOPEN_KEY: {
3397 __u8 key[TCP_FASTOPEN_KEY_BUF_LENGTH];
3398 __u8 *backup_key = NULL;
3399
3400 /* Allow a backup key as well to facilitate key rotation
3401 * First key is the active one.
3402 */
3403 if (optlen != TCP_FASTOPEN_KEY_LENGTH &&
3404 optlen != TCP_FASTOPEN_KEY_BUF_LENGTH)
3405 return -EINVAL;
3406
3407 if (copy_from_sockptr(key, optval, optlen))
3408 return -EFAULT;
3409
3410 if (optlen == TCP_FASTOPEN_KEY_BUF_LENGTH)
3411 backup_key = key + TCP_FASTOPEN_KEY_LENGTH;
3412
3413 return tcp_fastopen_reset_cipher(net, sk, key, backup_key);
3414 }
3415 default:
3416 /* fallthru */
3417 break;
3418 }
3419
3420 if (optlen < sizeof(int))
3421 return -EINVAL;
3422
3423 if (copy_from_sockptr(&val, optval, sizeof(val)))
3424 return -EFAULT;
3425
3426 lock_sock(sk);
3427
3428 switch (optname) {
3429 case TCP_MAXSEG:
3430 /* Values greater than interface MTU won't take effect. However
3431 * at the point when this call is done we typically don't yet
3432 * know which interface is going to be used
3433 */
3434 if (val && (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW)) {
3435 err = -EINVAL;
3436 break;
3437 }
3438 tp->rx_opt.user_mss = val;
3439 break;
3440
3441 case TCP_NODELAY:
3442 __tcp_sock_set_nodelay(sk, val);
3443 break;
3444
3445 case TCP_THIN_LINEAR_TIMEOUTS:
3446 if (val < 0 || val > 1)
3447 err = -EINVAL;
3448 else
3449 tp->thin_lto = val;
3450 break;
3451
3452 case TCP_THIN_DUPACK:
3453 if (val < 0 || val > 1)
3454 err = -EINVAL;
3455 break;
3456
3457 case TCP_REPAIR:
3458 if (!tcp_can_repair_sock(sk))
3459 err = -EPERM;
3460 else if (val == TCP_REPAIR_ON) {
3461 tp->repair = 1;
3462 sk->sk_reuse = SK_FORCE_REUSE;
3463 tp->repair_queue = TCP_NO_QUEUE;
3464 } else if (val == TCP_REPAIR_OFF) {
3465 tp->repair = 0;
3466 sk->sk_reuse = SK_NO_REUSE;
3467 tcp_send_window_probe(sk);
3468 } else if (val == TCP_REPAIR_OFF_NO_WP) {
3469 tp->repair = 0;
3470 sk->sk_reuse = SK_NO_REUSE;
3471 } else
3472 err = -EINVAL;
3473
3474 break;
3475
3476 case TCP_REPAIR_QUEUE:
3477 if (!tp->repair)
3478 err = -EPERM;
3479 else if ((unsigned int)val < TCP_QUEUES_NR)
3480 tp->repair_queue = val;
3481 else
3482 err = -EINVAL;
3483 break;
3484
3485 case TCP_QUEUE_SEQ:
3486 if (sk->sk_state != TCP_CLOSE) {
3487 err = -EPERM;
3488 } else if (tp->repair_queue == TCP_SEND_QUEUE) {
3489 if (!tcp_rtx_queue_empty(sk))
3490 err = -EPERM;
3491 else
3492 WRITE_ONCE(tp->write_seq, val);
3493 } else if (tp->repair_queue == TCP_RECV_QUEUE) {
3494 if (tp->rcv_nxt != tp->copied_seq) {
3495 err = -EPERM;
3496 } else {
3497 WRITE_ONCE(tp->rcv_nxt, val);
3498 WRITE_ONCE(tp->copied_seq, val);
3499 }
3500 } else {
3501 err = -EINVAL;
3502 }
3503 break;
3504
3505 case TCP_REPAIR_OPTIONS:
3506 if (!tp->repair)
3507 err = -EINVAL;
3508 else if (sk->sk_state == TCP_ESTABLISHED)
3509 err = tcp_repair_options_est(sk, optval, optlen);
3510 else
3511 err = -EPERM;
3512 break;
3513
3514 case TCP_CORK:
3515 __tcp_sock_set_cork(sk, val);
3516 break;
3517
3518 case TCP_KEEPIDLE:
3519 err = tcp_sock_set_keepidle_locked(sk, val);
3520 break;
3521 case TCP_KEEPINTVL:
3522 if (val < 1 || val > MAX_TCP_KEEPINTVL)
3523 err = -EINVAL;
3524 else
3525 tp->keepalive_intvl = val * HZ;
3526 break;
3527 case TCP_KEEPCNT:
3528 if (val < 1 || val > MAX_TCP_KEEPCNT)
3529 err = -EINVAL;
3530 else
3531 tp->keepalive_probes = val;
3532 break;
3533 case TCP_SYNCNT:
3534 if (val < 1 || val > MAX_TCP_SYNCNT)
3535 err = -EINVAL;
3536 else
3537 icsk->icsk_syn_retries = val;
3538 break;
3539
3540 case TCP_SAVE_SYN:
3541 /* 0: disable, 1: enable, 2: start from ether_header */
3542 if (val < 0 || val > 2)
3543 err = -EINVAL;
3544 else
3545 tp->save_syn = val;
3546 break;
3547
3548 case TCP_LINGER2:
3549 if (val < 0)
3550 tp->linger2 = -1;
3551 else if (val > TCP_FIN_TIMEOUT_MAX / HZ)
3552 tp->linger2 = TCP_FIN_TIMEOUT_MAX;
3553 else
3554 tp->linger2 = val * HZ;
3555 break;
3556
3557 case TCP_DEFER_ACCEPT:
3558 /* Translate value in seconds to number of retransmits */
3559 icsk->icsk_accept_queue.rskq_defer_accept =
3560 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
3561 TCP_RTO_MAX / HZ);
3562 break;
3563
3564 case TCP_WINDOW_CLAMP:
3565 err = tcp_set_window_clamp(sk, val);
3566 break;
3567
3568 case TCP_QUICKACK:
3569 __tcp_sock_set_quickack(sk, val);
3570 break;
3571
3572#ifdef CONFIG_TCP_MD5SIG
3573 case TCP_MD5SIG:
3574 case TCP_MD5SIG_EXT:
3575 err = tp->af_specific->md5_parse(sk, optname, optval, optlen);
3576 break;
3577#endif
3578 case TCP_USER_TIMEOUT:
3579 /* Cap the max time in ms TCP will retry or probe the window
3580 * before giving up and aborting (ETIMEDOUT) a connection.
3581 */
3582 if (val < 0)
3583 err = -EINVAL;
3584 else
3585 icsk->icsk_user_timeout = val;
3586 break;
3587
3588 case TCP_FASTOPEN:
3589 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
3590 TCPF_LISTEN))) {
3591 tcp_fastopen_init_key_once(net);
3592
3593 fastopen_queue_tune(sk, val);
3594 } else {
3595 err = -EINVAL;
3596 }
3597 break;
3598 case TCP_FASTOPEN_CONNECT:
3599 if (val > 1 || val < 0) {
3600 err = -EINVAL;
3601 } else if (net->ipv4.sysctl_tcp_fastopen & TFO_CLIENT_ENABLE) {
3602 if (sk->sk_state == TCP_CLOSE)
3603 tp->fastopen_connect = val;
3604 else
3605 err = -EINVAL;
3606 } else {
3607 err = -EOPNOTSUPP;
3608 }
3609 break;
3610 case TCP_FASTOPEN_NO_COOKIE:
3611 if (val > 1 || val < 0)
3612 err = -EINVAL;
3613 else if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3614 err = -EINVAL;
3615 else
3616 tp->fastopen_no_cookie = val;
3617 break;
3618 case TCP_TIMESTAMP:
3619 if (!tp->repair)
3620 err = -EPERM;
3621 else
3622 tp->tsoffset = val - tcp_time_stamp_raw();
3623 break;
3624 case TCP_REPAIR_WINDOW:
3625 err = tcp_repair_set_window(tp, optval, optlen);
3626 break;
3627 case TCP_NOTSENT_LOWAT:
3628 tp->notsent_lowat = val;
3629 sk->sk_write_space(sk);
3630 break;
3631 case TCP_INQ:
3632 if (val > 1 || val < 0)
3633 err = -EINVAL;
3634 else
3635 tp->recvmsg_inq = val;
3636 break;
3637 case TCP_TX_DELAY:
3638 if (val)
3639 tcp_enable_tx_delay();
3640 tp->tcp_tx_delay = val;
3641 break;
3642 default:
3643 err = -ENOPROTOOPT;
3644 break;
3645 }
3646
3647 release_sock(sk);
3648 return err;
3649}
3650
3651int tcp_setsockopt(struct sock *sk, int level, int optname, sockptr_t optval,
3652 unsigned int optlen)
3653{
3654 const struct inet_connection_sock *icsk = inet_csk(sk);
3655
3656 if (level != SOL_TCP)
3657 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
3658 optval, optlen);
3659 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
3660}
3661EXPORT_SYMBOL(tcp_setsockopt);
3662
3663static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
3664 struct tcp_info *info)
3665{
3666 u64 stats[__TCP_CHRONO_MAX], total = 0;
3667 enum tcp_chrono i;
3668
3669 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
3670 stats[i] = tp->chrono_stat[i - 1];
3671 if (i == tp->chrono_type)
3672 stats[i] += tcp_jiffies32 - tp->chrono_start;
3673 stats[i] *= USEC_PER_SEC / HZ;
3674 total += stats[i];
3675 }
3676
3677 info->tcpi_busy_time = total;
3678 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
3679 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
3680}
3681
3682/* Return information about state of tcp endpoint in API format. */
3683void tcp_get_info(struct sock *sk, struct tcp_info *info)
3684{
3685 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
3686 const struct inet_connection_sock *icsk = inet_csk(sk);
3687 unsigned long rate;
3688 u32 now;
3689 u64 rate64;
3690 bool slow;
3691
3692 memset(info, 0, sizeof(*info));
3693 if (sk->sk_type != SOCK_STREAM)
3694 return;
3695
3696 info->tcpi_state = inet_sk_state_load(sk);
3697
3698 /* Report meaningful fields for all TCP states, including listeners */
3699 rate = READ_ONCE(sk->sk_pacing_rate);
3700 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3701 info->tcpi_pacing_rate = rate64;
3702
3703 rate = READ_ONCE(sk->sk_max_pacing_rate);
3704 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3705 info->tcpi_max_pacing_rate = rate64;
3706
3707 info->tcpi_reordering = tp->reordering;
3708 info->tcpi_snd_cwnd = tp->snd_cwnd;
3709
3710 if (info->tcpi_state == TCP_LISTEN) {
3711 /* listeners aliased fields :
3712 * tcpi_unacked -> Number of children ready for accept()
3713 * tcpi_sacked -> max backlog
3714 */
3715 info->tcpi_unacked = READ_ONCE(sk->sk_ack_backlog);
3716 info->tcpi_sacked = READ_ONCE(sk->sk_max_ack_backlog);
3717 return;
3718 }
3719
3720 slow = lock_sock_fast(sk);
3721
3722 info->tcpi_ca_state = icsk->icsk_ca_state;
3723 info->tcpi_retransmits = icsk->icsk_retransmits;
3724 info->tcpi_probes = icsk->icsk_probes_out;
3725 info->tcpi_backoff = icsk->icsk_backoff;
3726
3727 if (tp->rx_opt.tstamp_ok)
3728 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
3729 if (tcp_is_sack(tp))
3730 info->tcpi_options |= TCPI_OPT_SACK;
3731 if (tp->rx_opt.wscale_ok) {
3732 info->tcpi_options |= TCPI_OPT_WSCALE;
3733 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
3734 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
3735 }
3736
3737 if (tp->ecn_flags & TCP_ECN_OK)
3738 info->tcpi_options |= TCPI_OPT_ECN;
3739 if (tp->ecn_flags & TCP_ECN_SEEN)
3740 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
3741 if (tp->syn_data_acked)
3742 info->tcpi_options |= TCPI_OPT_SYN_DATA;
3743
3744 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
3745 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
3746 info->tcpi_snd_mss = tp->mss_cache;
3747 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
3748
3749 info->tcpi_unacked = tp->packets_out;
3750 info->tcpi_sacked = tp->sacked_out;
3751
3752 info->tcpi_lost = tp->lost_out;
3753 info->tcpi_retrans = tp->retrans_out;
3754
3755 now = tcp_jiffies32;
3756 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
3757 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
3758 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
3759
3760 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
3761 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
3762 info->tcpi_rtt = tp->srtt_us >> 3;
3763 info->tcpi_rttvar = tp->mdev_us >> 2;
3764 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
3765 info->tcpi_advmss = tp->advmss;
3766
3767 info->tcpi_rcv_rtt = tp->rcv_rtt_est.rtt_us >> 3;
3768 info->tcpi_rcv_space = tp->rcvq_space.space;
3769
3770 info->tcpi_total_retrans = tp->total_retrans;
3771
3772 info->tcpi_bytes_acked = tp->bytes_acked;
3773 info->tcpi_bytes_received = tp->bytes_received;
3774 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
3775 tcp_get_info_chrono_stats(tp, info);
3776
3777 info->tcpi_segs_out = tp->segs_out;
3778 info->tcpi_segs_in = tp->segs_in;
3779
3780 info->tcpi_min_rtt = tcp_min_rtt(tp);
3781 info->tcpi_data_segs_in = tp->data_segs_in;
3782 info->tcpi_data_segs_out = tp->data_segs_out;
3783
3784 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
3785 rate64 = tcp_compute_delivery_rate(tp);
3786 if (rate64)
3787 info->tcpi_delivery_rate = rate64;
3788 info->tcpi_delivered = tp->delivered;
3789 info->tcpi_delivered_ce = tp->delivered_ce;
3790 info->tcpi_bytes_sent = tp->bytes_sent;
3791 info->tcpi_bytes_retrans = tp->bytes_retrans;
3792 info->tcpi_dsack_dups = tp->dsack_dups;
3793 info->tcpi_reord_seen = tp->reord_seen;
3794 info->tcpi_rcv_ooopack = tp->rcv_ooopack;
3795 info->tcpi_snd_wnd = tp->snd_wnd;
3796 info->tcpi_fastopen_client_fail = tp->fastopen_client_fail;
3797 unlock_sock_fast(sk, slow);
3798}
3799EXPORT_SYMBOL_GPL(tcp_get_info);
3800
3801static size_t tcp_opt_stats_get_size(void)
3802{
3803 return
3804 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BUSY */
3805 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_RWND_LIMITED */
3806 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_SNDBUF_LIMITED */
3807 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DATA_SEGS_OUT */
3808 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_TOTAL_RETRANS */
3809 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_PACING_RATE */
3810 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_DELIVERY_RATE */
3811 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_CWND */
3812 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORDERING */
3813 nla_total_size(sizeof(u32)) + /* TCP_NLA_MIN_RTT */
3814 nla_total_size(sizeof(u8)) + /* TCP_NLA_RECUR_RETRANS */
3815 nla_total_size(sizeof(u8)) + /* TCP_NLA_DELIVERY_RATE_APP_LMT */
3816 nla_total_size(sizeof(u32)) + /* TCP_NLA_SNDQ_SIZE */
3817 nla_total_size(sizeof(u8)) + /* TCP_NLA_CA_STATE */
3818 nla_total_size(sizeof(u32)) + /* TCP_NLA_SND_SSTHRESH */
3819 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED */
3820 nla_total_size(sizeof(u32)) + /* TCP_NLA_DELIVERED_CE */
3821 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_SENT */
3822 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_BYTES_RETRANS */
3823 nla_total_size(sizeof(u32)) + /* TCP_NLA_DSACK_DUPS */
3824 nla_total_size(sizeof(u32)) + /* TCP_NLA_REORD_SEEN */
3825 nla_total_size(sizeof(u32)) + /* TCP_NLA_SRTT */
3826 nla_total_size(sizeof(u16)) + /* TCP_NLA_TIMEOUT_REHASH */
3827 nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */
3828 nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */
3829 nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */
3830 0;
3831}
3832
3833/* Returns TTL or hop limit of an incoming packet from skb. */
3834static u8 tcp_skb_ttl_or_hop_limit(const struct sk_buff *skb)
3835{
3836 if (skb->protocol == htons(ETH_P_IP))
3837 return ip_hdr(skb)->ttl;
3838 else if (skb->protocol == htons(ETH_P_IPV6))
3839 return ipv6_hdr(skb)->hop_limit;
3840 else
3841 return 0;
3842}
3843
3844struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk,
3845 const struct sk_buff *orig_skb,
3846 const struct sk_buff *ack_skb)
3847{
3848 const struct tcp_sock *tp = tcp_sk(sk);
3849 struct sk_buff *stats;
3850 struct tcp_info info;
3851 unsigned long rate;
3852 u64 rate64;
3853
3854 stats = alloc_skb(tcp_opt_stats_get_size(), GFP_ATOMIC);
3855 if (!stats)
3856 return NULL;
3857
3858 tcp_get_info_chrono_stats(tp, &info);
3859 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
3860 info.tcpi_busy_time, TCP_NLA_PAD);
3861 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
3862 info.tcpi_rwnd_limited, TCP_NLA_PAD);
3863 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
3864 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
3865 nla_put_u64_64bit(stats, TCP_NLA_DATA_SEGS_OUT,
3866 tp->data_segs_out, TCP_NLA_PAD);
3867 nla_put_u64_64bit(stats, TCP_NLA_TOTAL_RETRANS,
3868 tp->total_retrans, TCP_NLA_PAD);
3869
3870 rate = READ_ONCE(sk->sk_pacing_rate);
3871 rate64 = (rate != ~0UL) ? rate : ~0ULL;
3872 nla_put_u64_64bit(stats, TCP_NLA_PACING_RATE, rate64, TCP_NLA_PAD);
3873
3874 rate64 = tcp_compute_delivery_rate(tp);
3875 nla_put_u64_64bit(stats, TCP_NLA_DELIVERY_RATE, rate64, TCP_NLA_PAD);
3876
3877 nla_put_u32(stats, TCP_NLA_SND_CWND, tp->snd_cwnd);
3878 nla_put_u32(stats, TCP_NLA_REORDERING, tp->reordering);
3879 nla_put_u32(stats, TCP_NLA_MIN_RTT, tcp_min_rtt(tp));
3880
3881 nla_put_u8(stats, TCP_NLA_RECUR_RETRANS, inet_csk(sk)->icsk_retransmits);
3882 nla_put_u8(stats, TCP_NLA_DELIVERY_RATE_APP_LMT, !!tp->rate_app_limited);
3883 nla_put_u32(stats, TCP_NLA_SND_SSTHRESH, tp->snd_ssthresh);
3884 nla_put_u32(stats, TCP_NLA_DELIVERED, tp->delivered);
3885 nla_put_u32(stats, TCP_NLA_DELIVERED_CE, tp->delivered_ce);
3886
3887 nla_put_u32(stats, TCP_NLA_SNDQ_SIZE, tp->write_seq - tp->snd_una);
3888 nla_put_u8(stats, TCP_NLA_CA_STATE, inet_csk(sk)->icsk_ca_state);
3889
3890 nla_put_u64_64bit(stats, TCP_NLA_BYTES_SENT, tp->bytes_sent,
3891 TCP_NLA_PAD);
3892 nla_put_u64_64bit(stats, TCP_NLA_BYTES_RETRANS, tp->bytes_retrans,
3893 TCP_NLA_PAD);
3894 nla_put_u32(stats, TCP_NLA_DSACK_DUPS, tp->dsack_dups);
3895 nla_put_u32(stats, TCP_NLA_REORD_SEEN, tp->reord_seen);
3896 nla_put_u32(stats, TCP_NLA_SRTT, tp->srtt_us >> 3);
3897 nla_put_u16(stats, TCP_NLA_TIMEOUT_REHASH, tp->timeout_rehash);
3898 nla_put_u32(stats, TCP_NLA_BYTES_NOTSENT,
3899 max_t(int, 0, tp->write_seq - tp->snd_nxt));
3900 nla_put_u64_64bit(stats, TCP_NLA_EDT, orig_skb->skb_mstamp_ns,
3901 TCP_NLA_PAD);
3902 if (ack_skb)
3903 nla_put_u8(stats, TCP_NLA_TTL,
3904 tcp_skb_ttl_or_hop_limit(ack_skb));
3905
3906 return stats;
3907}
3908
3909static int do_tcp_getsockopt(struct sock *sk, int level,
3910 int optname, char __user *optval, int __user *optlen)
3911{
3912 struct inet_connection_sock *icsk = inet_csk(sk);
3913 struct tcp_sock *tp = tcp_sk(sk);
3914 struct net *net = sock_net(sk);
3915 int val, len;
3916
3917 if (get_user(len, optlen))
3918 return -EFAULT;
3919
3920 len = min_t(unsigned int, len, sizeof(int));
3921
3922 if (len < 0)
3923 return -EINVAL;
3924
3925 switch (optname) {
3926 case TCP_MAXSEG:
3927 val = tp->mss_cache;
3928 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
3929 val = tp->rx_opt.user_mss;
3930 if (tp->repair)
3931 val = tp->rx_opt.mss_clamp;
3932 break;
3933 case TCP_NODELAY:
3934 val = !!(tp->nonagle&TCP_NAGLE_OFF);
3935 break;
3936 case TCP_CORK:
3937 val = !!(tp->nonagle&TCP_NAGLE_CORK);
3938 break;
3939 case TCP_KEEPIDLE:
3940 val = keepalive_time_when(tp) / HZ;
3941 break;
3942 case TCP_KEEPINTVL:
3943 val = keepalive_intvl_when(tp) / HZ;
3944 break;
3945 case TCP_KEEPCNT:
3946 val = keepalive_probes(tp);
3947 break;
3948 case TCP_SYNCNT:
3949 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
3950 break;
3951 case TCP_LINGER2:
3952 val = tp->linger2;
3953 if (val >= 0)
3954 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
3955 break;
3956 case TCP_DEFER_ACCEPT:
3957 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
3958 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
3959 break;
3960 case TCP_WINDOW_CLAMP:
3961 val = tp->window_clamp;
3962 break;
3963 case TCP_INFO: {
3964 struct tcp_info info;
3965
3966 if (get_user(len, optlen))
3967 return -EFAULT;
3968
3969 tcp_get_info(sk, &info);
3970
3971 len = min_t(unsigned int, len, sizeof(info));
3972 if (put_user(len, optlen))
3973 return -EFAULT;
3974 if (copy_to_user(optval, &info, len))
3975 return -EFAULT;
3976 return 0;
3977 }
3978 case TCP_CC_INFO: {
3979 const struct tcp_congestion_ops *ca_ops;
3980 union tcp_cc_info info;
3981 size_t sz = 0;
3982 int attr;
3983
3984 if (get_user(len, optlen))
3985 return -EFAULT;
3986
3987 ca_ops = icsk->icsk_ca_ops;
3988 if (ca_ops && ca_ops->get_info)
3989 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
3990
3991 len = min_t(unsigned int, len, sz);
3992 if (put_user(len, optlen))
3993 return -EFAULT;
3994 if (copy_to_user(optval, &info, len))
3995 return -EFAULT;
3996 return 0;
3997 }
3998 case TCP_QUICKACK:
3999 val = !inet_csk_in_pingpong_mode(sk);
4000 break;
4001
4002 case TCP_CONGESTION:
4003 if (get_user(len, optlen))
4004 return -EFAULT;
4005 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
4006 if (put_user(len, optlen))
4007 return -EFAULT;
4008 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
4009 return -EFAULT;
4010 return 0;
4011
4012 case TCP_ULP:
4013 if (get_user(len, optlen))
4014 return -EFAULT;
4015 len = min_t(unsigned int, len, TCP_ULP_NAME_MAX);
4016 if (!icsk->icsk_ulp_ops) {
4017 if (put_user(0, optlen))
4018 return -EFAULT;
4019 return 0;
4020 }
4021 if (put_user(len, optlen))
4022 return -EFAULT;
4023 if (copy_to_user(optval, icsk->icsk_ulp_ops->name, len))
4024 return -EFAULT;
4025 return 0;
4026
4027 case TCP_FASTOPEN_KEY: {
4028 u64 key[TCP_FASTOPEN_KEY_BUF_LENGTH / sizeof(u64)];
4029 unsigned int key_len;
4030
4031 if (get_user(len, optlen))
4032 return -EFAULT;
4033
4034 key_len = tcp_fastopen_get_cipher(net, icsk, key) *
4035 TCP_FASTOPEN_KEY_LENGTH;
4036 len = min_t(unsigned int, len, key_len);
4037 if (put_user(len, optlen))
4038 return -EFAULT;
4039 if (copy_to_user(optval, key, len))
4040 return -EFAULT;
4041 return 0;
4042 }
4043 case TCP_THIN_LINEAR_TIMEOUTS:
4044 val = tp->thin_lto;
4045 break;
4046
4047 case TCP_THIN_DUPACK:
4048 val = 0;
4049 break;
4050
4051 case TCP_REPAIR:
4052 val = tp->repair;
4053 break;
4054
4055 case TCP_REPAIR_QUEUE:
4056 if (tp->repair)
4057 val = tp->repair_queue;
4058 else
4059 return -EINVAL;
4060 break;
4061
4062 case TCP_REPAIR_WINDOW: {
4063 struct tcp_repair_window opt;
4064
4065 if (get_user(len, optlen))
4066 return -EFAULT;
4067
4068 if (len != sizeof(opt))
4069 return -EINVAL;
4070
4071 if (!tp->repair)
4072 return -EPERM;
4073
4074 opt.snd_wl1 = tp->snd_wl1;
4075 opt.snd_wnd = tp->snd_wnd;
4076 opt.max_window = tp->max_window;
4077 opt.rcv_wnd = tp->rcv_wnd;
4078 opt.rcv_wup = tp->rcv_wup;
4079
4080 if (copy_to_user(optval, &opt, len))
4081 return -EFAULT;
4082 return 0;
4083 }
4084 case TCP_QUEUE_SEQ:
4085 if (tp->repair_queue == TCP_SEND_QUEUE)
4086 val = tp->write_seq;
4087 else if (tp->repair_queue == TCP_RECV_QUEUE)
4088 val = tp->rcv_nxt;
4089 else
4090 return -EINVAL;
4091 break;
4092
4093 case TCP_USER_TIMEOUT:
4094 val = icsk->icsk_user_timeout;
4095 break;
4096
4097 case TCP_FASTOPEN:
4098 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
4099 break;
4100
4101 case TCP_FASTOPEN_CONNECT:
4102 val = tp->fastopen_connect;
4103 break;
4104
4105 case TCP_FASTOPEN_NO_COOKIE:
4106 val = tp->fastopen_no_cookie;
4107 break;
4108
4109 case TCP_TX_DELAY:
4110 val = tp->tcp_tx_delay;
4111 break;
4112
4113 case TCP_TIMESTAMP:
4114 val = tcp_time_stamp_raw() + tp->tsoffset;
4115 break;
4116 case TCP_NOTSENT_LOWAT:
4117 val = tp->notsent_lowat;
4118 break;
4119 case TCP_INQ:
4120 val = tp->recvmsg_inq;
4121 break;
4122 case TCP_SAVE_SYN:
4123 val = tp->save_syn;
4124 break;
4125 case TCP_SAVED_SYN: {
4126 if (get_user(len, optlen))
4127 return -EFAULT;
4128
4129 lock_sock(sk);
4130 if (tp->saved_syn) {
4131 if (len < tcp_saved_syn_len(tp->saved_syn)) {
4132 if (put_user(tcp_saved_syn_len(tp->saved_syn),
4133 optlen)) {
4134 release_sock(sk);
4135 return -EFAULT;
4136 }
4137 release_sock(sk);
4138 return -EINVAL;
4139 }
4140 len = tcp_saved_syn_len(tp->saved_syn);
4141 if (put_user(len, optlen)) {
4142 release_sock(sk);
4143 return -EFAULT;
4144 }
4145 if (copy_to_user(optval, tp->saved_syn->data, len)) {
4146 release_sock(sk);
4147 return -EFAULT;
4148 }
4149 tcp_saved_syn_free(tp);
4150 release_sock(sk);
4151 } else {
4152 release_sock(sk);
4153 len = 0;
4154 if (put_user(len, optlen))
4155 return -EFAULT;
4156 }
4157 return 0;
4158 }
4159#ifdef CONFIG_MMU
4160 case TCP_ZEROCOPY_RECEIVE: {
4161 struct scm_timestamping_internal tss;
4162 struct tcp_zerocopy_receive zc = {};
4163 int err;
4164
4165 if (get_user(len, optlen))
4166 return -EFAULT;
4167 if (len < 0 ||
4168 len < offsetofend(struct tcp_zerocopy_receive, length))
4169 return -EINVAL;
4170 if (unlikely(len > sizeof(zc))) {
4171 err = check_zeroed_user(optval + sizeof(zc),
4172 len - sizeof(zc));
4173 if (err < 1)
4174 return err == 0 ? -EINVAL : err;
4175 len = sizeof(zc);
4176 if (put_user(len, optlen))
4177 return -EFAULT;
4178 }
4179 if (copy_from_user(&zc, optval, len))
4180 return -EFAULT;
4181 if (zc.reserved)
4182 return -EINVAL;
4183 if (zc.msg_flags & ~(TCP_VALID_ZC_MSG_FLAGS))
4184 return -EINVAL;
4185 lock_sock(sk);
4186 err = tcp_zerocopy_receive(sk, &zc, &tss);
4187 err = BPF_CGROUP_RUN_PROG_GETSOCKOPT_KERN(sk, level, optname,
4188 &zc, &len, err);
4189 release_sock(sk);
4190 if (len >= offsetofend(struct tcp_zerocopy_receive, msg_flags))
4191 goto zerocopy_rcv_cmsg;
4192 switch (len) {
4193 case offsetofend(struct tcp_zerocopy_receive, msg_flags):
4194 goto zerocopy_rcv_cmsg;
4195 case offsetofend(struct tcp_zerocopy_receive, msg_controllen):
4196 case offsetofend(struct tcp_zerocopy_receive, msg_control):
4197 case offsetofend(struct tcp_zerocopy_receive, flags):
4198 case offsetofend(struct tcp_zerocopy_receive, copybuf_len):
4199 case offsetofend(struct tcp_zerocopy_receive, copybuf_address):
4200 case offsetofend(struct tcp_zerocopy_receive, err):
4201 goto zerocopy_rcv_sk_err;
4202 case offsetofend(struct tcp_zerocopy_receive, inq):
4203 goto zerocopy_rcv_inq;
4204 case offsetofend(struct tcp_zerocopy_receive, length):
4205 default:
4206 goto zerocopy_rcv_out;
4207 }
4208zerocopy_rcv_cmsg:
4209 if (zc.msg_flags & TCP_CMSG_TS)
4210 tcp_zc_finalize_rx_tstamp(sk, &zc, &tss);
4211 else
4212 zc.msg_flags = 0;
4213zerocopy_rcv_sk_err:
4214 if (!err)
4215 zc.err = sock_error(sk);
4216zerocopy_rcv_inq:
4217 zc.inq = tcp_inq_hint(sk);
4218zerocopy_rcv_out:
4219 if (!err && copy_to_user(optval, &zc, len))
4220 err = -EFAULT;
4221 return err;
4222 }
4223#endif
4224 default:
4225 return -ENOPROTOOPT;
4226 }
4227
4228 if (put_user(len, optlen))
4229 return -EFAULT;
4230 if (copy_to_user(optval, &val, len))
4231 return -EFAULT;
4232 return 0;
4233}
4234
4235bool tcp_bpf_bypass_getsockopt(int level, int optname)
4236{
4237 /* TCP do_tcp_getsockopt has optimized getsockopt implementation
4238 * to avoid extra socket lock for TCP_ZEROCOPY_RECEIVE.
4239 */
4240 if (level == SOL_TCP && optname == TCP_ZEROCOPY_RECEIVE)
4241 return true;
4242
4243 return false;
4244}
4245EXPORT_SYMBOL(tcp_bpf_bypass_getsockopt);
4246
4247int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
4248 int __user *optlen)
4249{
4250 struct inet_connection_sock *icsk = inet_csk(sk);
4251
4252 if (level != SOL_TCP)
4253 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
4254 optval, optlen);
4255 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
4256}
4257EXPORT_SYMBOL(tcp_getsockopt);
4258
4259#ifdef CONFIG_TCP_MD5SIG
4260static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
4261static DEFINE_MUTEX(tcp_md5sig_mutex);
4262static bool tcp_md5sig_pool_populated = false;
4263
4264static void __tcp_alloc_md5sig_pool(void)
4265{
4266 struct crypto_ahash *hash;
4267 int cpu;
4268
4269 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
4270 if (IS_ERR(hash))
4271 return;
4272
4273 for_each_possible_cpu(cpu) {
4274 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
4275 struct ahash_request *req;
4276
4277 if (!scratch) {
4278 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
4279 sizeof(struct tcphdr),
4280 GFP_KERNEL,
4281 cpu_to_node(cpu));
4282 if (!scratch)
4283 return;
4284 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
4285 }
4286 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
4287 continue;
4288
4289 req = ahash_request_alloc(hash, GFP_KERNEL);
4290 if (!req)
4291 return;
4292
4293 ahash_request_set_callback(req, 0, NULL, NULL);
4294
4295 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
4296 }
4297 /* before setting tcp_md5sig_pool_populated, we must commit all writes
4298 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
4299 */
4300 smp_wmb();
4301 tcp_md5sig_pool_populated = true;
4302}
4303
4304bool tcp_alloc_md5sig_pool(void)
4305{
4306 if (unlikely(!tcp_md5sig_pool_populated)) {
4307 mutex_lock(&tcp_md5sig_mutex);
4308
4309 if (!tcp_md5sig_pool_populated) {
4310 __tcp_alloc_md5sig_pool();
4311 if (tcp_md5sig_pool_populated)
4312 static_branch_inc(&tcp_md5_needed);
4313 }
4314
4315 mutex_unlock(&tcp_md5sig_mutex);
4316 }
4317 return tcp_md5sig_pool_populated;
4318}
4319EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
4320
4321
4322/**
4323 * tcp_get_md5sig_pool - get md5sig_pool for this user
4324 *
4325 * We use percpu structure, so if we succeed, we exit with preemption
4326 * and BH disabled, to make sure another thread or softirq handling
4327 * wont try to get same context.
4328 */
4329struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
4330{
4331 local_bh_disable();
4332
4333 if (tcp_md5sig_pool_populated) {
4334 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
4335 smp_rmb();
4336 return this_cpu_ptr(&tcp_md5sig_pool);
4337 }
4338 local_bh_enable();
4339 return NULL;
4340}
4341EXPORT_SYMBOL(tcp_get_md5sig_pool);
4342
4343int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
4344 const struct sk_buff *skb, unsigned int header_len)
4345{
4346 struct scatterlist sg;
4347 const struct tcphdr *tp = tcp_hdr(skb);
4348 struct ahash_request *req = hp->md5_req;
4349 unsigned int i;
4350 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
4351 skb_headlen(skb) - header_len : 0;
4352 const struct skb_shared_info *shi = skb_shinfo(skb);
4353 struct sk_buff *frag_iter;
4354
4355 sg_init_table(&sg, 1);
4356
4357 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
4358 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
4359 if (crypto_ahash_update(req))
4360 return 1;
4361
4362 for (i = 0; i < shi->nr_frags; ++i) {
4363 const skb_frag_t *f = &shi->frags[i];
4364 unsigned int offset = skb_frag_off(f);
4365 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
4366
4367 sg_set_page(&sg, page, skb_frag_size(f),
4368 offset_in_page(offset));
4369 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
4370 if (crypto_ahash_update(req))
4371 return 1;
4372 }
4373
4374 skb_walk_frags(skb, frag_iter)
4375 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
4376 return 1;
4377
4378 return 0;
4379}
4380EXPORT_SYMBOL(tcp_md5_hash_skb_data);
4381
4382int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
4383{
4384 u8 keylen = READ_ONCE(key->keylen); /* paired with WRITE_ONCE() in tcp_md5_do_add */
4385 struct scatterlist sg;
4386
4387 sg_init_one(&sg, key->key, keylen);
4388 ahash_request_set_crypt(hp->md5_req, &sg, NULL, keylen);
4389
4390 /* We use data_race() because tcp_md5_do_add() might change key->key under us */
4391 return data_race(crypto_ahash_update(hp->md5_req));
4392}
4393EXPORT_SYMBOL(tcp_md5_hash_key);
4394
4395#endif
4396
4397void tcp_done(struct sock *sk)
4398{
4399 struct request_sock *req;
4400
4401 /* We might be called with a new socket, after
4402 * inet_csk_prepare_forced_close() has been called
4403 * so we can not use lockdep_sock_is_held(sk)
4404 */
4405 req = rcu_dereference_protected(tcp_sk(sk)->fastopen_rsk, 1);
4406
4407 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
4408 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
4409
4410 tcp_set_state(sk, TCP_CLOSE);
4411 tcp_clear_xmit_timers(sk);
4412 if (req)
4413 reqsk_fastopen_remove(sk, req, false);
4414
4415 sk->sk_shutdown = SHUTDOWN_MASK;
4416
4417 if (!sock_flag(sk, SOCK_DEAD))
4418 sk->sk_state_change(sk);
4419 else
4420 inet_csk_destroy_sock(sk);
4421}
4422EXPORT_SYMBOL_GPL(tcp_done);
4423
4424int tcp_abort(struct sock *sk, int err)
4425{
4426 if (!sk_fullsock(sk)) {
4427 if (sk->sk_state == TCP_NEW_SYN_RECV) {
4428 struct request_sock *req = inet_reqsk(sk);
4429
4430 local_bh_disable();
4431 inet_csk_reqsk_queue_drop(req->rsk_listener, req);
4432 local_bh_enable();
4433 return 0;
4434 }
4435 return -EOPNOTSUPP;
4436 }
4437
4438 /* Don't race with userspace socket closes such as tcp_close. */
4439 lock_sock(sk);
4440
4441 if (sk->sk_state == TCP_LISTEN) {
4442 tcp_set_state(sk, TCP_CLOSE);
4443 inet_csk_listen_stop(sk);
4444 }
4445
4446 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
4447 local_bh_disable();
4448 bh_lock_sock(sk);
4449
4450 if (!sock_flag(sk, SOCK_DEAD)) {
4451 sk->sk_err = err;
4452 /* This barrier is coupled with smp_rmb() in tcp_poll() */
4453 smp_wmb();
4454 sk_error_report(sk);
4455 if (tcp_need_reset(sk->sk_state))
4456 tcp_send_active_reset(sk, GFP_ATOMIC);
4457 tcp_done(sk);
4458 }
4459
4460 bh_unlock_sock(sk);
4461 local_bh_enable();
4462 tcp_write_queue_purge(sk);
4463 release_sock(sk);
4464 return 0;
4465}
4466EXPORT_SYMBOL_GPL(tcp_abort);
4467
4468extern struct tcp_congestion_ops tcp_reno;
4469
4470static __initdata unsigned long thash_entries;
4471static int __init set_thash_entries(char *str)
4472{
4473 ssize_t ret;
4474
4475 if (!str)
4476 return 0;
4477
4478 ret = kstrtoul(str, 0, &thash_entries);
4479 if (ret)
4480 return 0;
4481
4482 return 1;
4483}
4484__setup("thash_entries=", set_thash_entries);
4485
4486static void __init tcp_init_mem(void)
4487{
4488 unsigned long limit = nr_free_buffer_pages() / 16;
4489
4490 limit = max(limit, 128UL);
4491 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
4492 sysctl_tcp_mem[1] = limit; /* 6.25 % */
4493 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
4494}
4495
4496void __init tcp_init(void)
4497{
4498 int max_rshare, max_wshare, cnt;
4499 unsigned long limit;
4500 unsigned int i;
4501
4502 BUILD_BUG_ON(TCP_MIN_SND_MSS <= MAX_TCP_OPTION_SPACE);
4503 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
4504 sizeof_field(struct sk_buff, cb));
4505
4506 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
4507 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
4508 inet_hashinfo_init(&tcp_hashinfo);
4509 inet_hashinfo2_init(&tcp_hashinfo, "tcp_listen_portaddr_hash",
4510 thash_entries, 21, /* one slot per 2 MB*/
4511 0, 64 * 1024);
4512 tcp_hashinfo.bind_bucket_cachep =
4513 kmem_cache_create("tcp_bind_bucket",
4514 sizeof(struct inet_bind_bucket), 0,
4515 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
4516
4517 /* Size and allocate the main established and bind bucket
4518 * hash tables.
4519 *
4520 * The methodology is similar to that of the buffer cache.
4521 */
4522 tcp_hashinfo.ehash =
4523 alloc_large_system_hash("TCP established",
4524 sizeof(struct inet_ehash_bucket),
4525 thash_entries,
4526 17, /* one slot per 128 KB of memory */
4527 0,
4528 NULL,
4529 &tcp_hashinfo.ehash_mask,
4530 0,
4531 thash_entries ? 0 : 512 * 1024);
4532 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
4533 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
4534
4535 if (inet_ehash_locks_alloc(&tcp_hashinfo))
4536 panic("TCP: failed to alloc ehash_locks");
4537 tcp_hashinfo.bhash =
4538 alloc_large_system_hash("TCP bind",
4539 sizeof(struct inet_bind_hashbucket),
4540 tcp_hashinfo.ehash_mask + 1,
4541 17, /* one slot per 128 KB of memory */
4542 0,
4543 &tcp_hashinfo.bhash_size,
4544 NULL,
4545 0,
4546 64 * 1024);
4547 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
4548 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
4549 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
4550 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
4551 }
4552
4553
4554 cnt = tcp_hashinfo.ehash_mask + 1;
4555 sysctl_tcp_max_orphans = cnt / 2;
4556
4557 tcp_init_mem();
4558 /* Set per-socket limits to no more than 1/128 the pressure threshold */
4559 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
4560 max_wshare = min(4UL*1024*1024, limit);
4561 max_rshare = min(6UL*1024*1024, limit);
4562
4563 init_net.ipv4.sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
4564 init_net.ipv4.sysctl_tcp_wmem[1] = 16*1024;
4565 init_net.ipv4.sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
4566
4567 init_net.ipv4.sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
4568 init_net.ipv4.sysctl_tcp_rmem[1] = 131072;
4569 init_net.ipv4.sysctl_tcp_rmem[2] = max(131072, max_rshare);
4570
4571 pr_info("Hash tables configured (established %u bind %u)\n",
4572 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
4573
4574 tcp_v4_init();
4575 tcp_metrics_init();
4576 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
4577 tcp_tasklet_init();
4578 mptcp_init();
4579}
1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 *
20 * Fixes:
21 * Alan Cox : Numerous verify_area() calls
22 * Alan Cox : Set the ACK bit on a reset
23 * Alan Cox : Stopped it crashing if it closed while
24 * sk->inuse=1 and was trying to connect
25 * (tcp_err()).
26 * Alan Cox : All icmp error handling was broken
27 * pointers passed where wrong and the
28 * socket was looked up backwards. Nobody
29 * tested any icmp error code obviously.
30 * Alan Cox : tcp_err() now handled properly. It
31 * wakes people on errors. poll
32 * behaves and the icmp error race
33 * has gone by moving it into sock.c
34 * Alan Cox : tcp_send_reset() fixed to work for
35 * everything not just packets for
36 * unknown sockets.
37 * Alan Cox : tcp option processing.
38 * Alan Cox : Reset tweaked (still not 100%) [Had
39 * syn rule wrong]
40 * Herp Rosmanith : More reset fixes
41 * Alan Cox : No longer acks invalid rst frames.
42 * Acking any kind of RST is right out.
43 * Alan Cox : Sets an ignore me flag on an rst
44 * receive otherwise odd bits of prattle
45 * escape still
46 * Alan Cox : Fixed another acking RST frame bug.
47 * Should stop LAN workplace lockups.
48 * Alan Cox : Some tidyups using the new skb list
49 * facilities
50 * Alan Cox : sk->keepopen now seems to work
51 * Alan Cox : Pulls options out correctly on accepts
52 * Alan Cox : Fixed assorted sk->rqueue->next errors
53 * Alan Cox : PSH doesn't end a TCP read. Switched a
54 * bit to skb ops.
55 * Alan Cox : Tidied tcp_data to avoid a potential
56 * nasty.
57 * Alan Cox : Added some better commenting, as the
58 * tcp is hard to follow
59 * Alan Cox : Removed incorrect check for 20 * psh
60 * Michael O'Reilly : ack < copied bug fix.
61 * Johannes Stille : Misc tcp fixes (not all in yet).
62 * Alan Cox : FIN with no memory -> CRASH
63 * Alan Cox : Added socket option proto entries.
64 * Also added awareness of them to accept.
65 * Alan Cox : Added TCP options (SOL_TCP)
66 * Alan Cox : Switched wakeup calls to callbacks,
67 * so the kernel can layer network
68 * sockets.
69 * Alan Cox : Use ip_tos/ip_ttl settings.
70 * Alan Cox : Handle FIN (more) properly (we hope).
71 * Alan Cox : RST frames sent on unsynchronised
72 * state ack error.
73 * Alan Cox : Put in missing check for SYN bit.
74 * Alan Cox : Added tcp_select_window() aka NET2E
75 * window non shrink trick.
76 * Alan Cox : Added a couple of small NET2E timer
77 * fixes
78 * Charles Hedrick : TCP fixes
79 * Toomas Tamm : TCP window fixes
80 * Alan Cox : Small URG fix to rlogin ^C ack fight
81 * Charles Hedrick : Rewrote most of it to actually work
82 * Linus : Rewrote tcp_read() and URG handling
83 * completely
84 * Gerhard Koerting: Fixed some missing timer handling
85 * Matthew Dillon : Reworked TCP machine states as per RFC
86 * Gerhard Koerting: PC/TCP workarounds
87 * Adam Caldwell : Assorted timer/timing errors
88 * Matthew Dillon : Fixed another RST bug
89 * Alan Cox : Move to kernel side addressing changes.
90 * Alan Cox : Beginning work on TCP fastpathing
91 * (not yet usable)
92 * Arnt Gulbrandsen: Turbocharged tcp_check() routine.
93 * Alan Cox : TCP fast path debugging
94 * Alan Cox : Window clamping
95 * Michael Riepe : Bug in tcp_check()
96 * Matt Dillon : More TCP improvements and RST bug fixes
97 * Matt Dillon : Yet more small nasties remove from the
98 * TCP code (Be very nice to this man if
99 * tcp finally works 100%) 8)
100 * Alan Cox : BSD accept semantics.
101 * Alan Cox : Reset on closedown bug.
102 * Peter De Schrijver : ENOTCONN check missing in tcp_sendto().
103 * Michael Pall : Handle poll() after URG properly in
104 * all cases.
105 * Michael Pall : Undo the last fix in tcp_read_urg()
106 * (multi URG PUSH broke rlogin).
107 * Michael Pall : Fix the multi URG PUSH problem in
108 * tcp_readable(), poll() after URG
109 * works now.
110 * Michael Pall : recv(...,MSG_OOB) never blocks in the
111 * BSD api.
112 * Alan Cox : Changed the semantics of sk->socket to
113 * fix a race and a signal problem with
114 * accept() and async I/O.
115 * Alan Cox : Relaxed the rules on tcp_sendto().
116 * Yury Shevchuk : Really fixed accept() blocking problem.
117 * Craig I. Hagan : Allow for BSD compatible TIME_WAIT for
118 * clients/servers which listen in on
119 * fixed ports.
120 * Alan Cox : Cleaned the above up and shrank it to
121 * a sensible code size.
122 * Alan Cox : Self connect lockup fix.
123 * Alan Cox : No connect to multicast.
124 * Ross Biro : Close unaccepted children on master
125 * socket close.
126 * Alan Cox : Reset tracing code.
127 * Alan Cox : Spurious resets on shutdown.
128 * Alan Cox : Giant 15 minute/60 second timer error
129 * Alan Cox : Small whoops in polling before an
130 * accept.
131 * Alan Cox : Kept the state trace facility since
132 * it's handy for debugging.
133 * Alan Cox : More reset handler fixes.
134 * Alan Cox : Started rewriting the code based on
135 * the RFC's for other useful protocol
136 * references see: Comer, KA9Q NOS, and
137 * for a reference on the difference
138 * between specifications and how BSD
139 * works see the 4.4lite source.
140 * A.N.Kuznetsov : Don't time wait on completion of tidy
141 * close.
142 * Linus Torvalds : Fin/Shutdown & copied_seq changes.
143 * Linus Torvalds : Fixed BSD port reuse to work first syn
144 * Alan Cox : Reimplemented timers as per the RFC
145 * and using multiple timers for sanity.
146 * Alan Cox : Small bug fixes, and a lot of new
147 * comments.
148 * Alan Cox : Fixed dual reader crash by locking
149 * the buffers (much like datagram.c)
150 * Alan Cox : Fixed stuck sockets in probe. A probe
151 * now gets fed up of retrying without
152 * (even a no space) answer.
153 * Alan Cox : Extracted closing code better
154 * Alan Cox : Fixed the closing state machine to
155 * resemble the RFC.
156 * Alan Cox : More 'per spec' fixes.
157 * Jorge Cwik : Even faster checksumming.
158 * Alan Cox : tcp_data() doesn't ack illegal PSH
159 * only frames. At least one pc tcp stack
160 * generates them.
161 * Alan Cox : Cache last socket.
162 * Alan Cox : Per route irtt.
163 * Matt Day : poll()->select() match BSD precisely on error
164 * Alan Cox : New buffers
165 * Marc Tamsky : Various sk->prot->retransmits and
166 * sk->retransmits misupdating fixed.
167 * Fixed tcp_write_timeout: stuck close,
168 * and TCP syn retries gets used now.
169 * Mark Yarvis : In tcp_read_wakeup(), don't send an
170 * ack if state is TCP_CLOSED.
171 * Alan Cox : Look up device on a retransmit - routes may
172 * change. Doesn't yet cope with MSS shrink right
173 * but it's a start!
174 * Marc Tamsky : Closing in closing fixes.
175 * Mike Shaver : RFC1122 verifications.
176 * Alan Cox : rcv_saddr errors.
177 * Alan Cox : Block double connect().
178 * Alan Cox : Small hooks for enSKIP.
179 * Alexey Kuznetsov: Path MTU discovery.
180 * Alan Cox : Support soft errors.
181 * Alan Cox : Fix MTU discovery pathological case
182 * when the remote claims no mtu!
183 * Marc Tamsky : TCP_CLOSE fix.
184 * Colin (G3TNE) : Send a reset on syn ack replies in
185 * window but wrong (fixes NT lpd problems)
186 * Pedro Roque : Better TCP window handling, delayed ack.
187 * Joerg Reuter : No modification of locked buffers in
188 * tcp_do_retransmit()
189 * Eric Schenk : Changed receiver side silly window
190 * avoidance algorithm to BSD style
191 * algorithm. This doubles throughput
192 * against machines running Solaris,
193 * and seems to result in general
194 * improvement.
195 * Stefan Magdalinski : adjusted tcp_readable() to fix FIONREAD
196 * Willy Konynenberg : Transparent proxying support.
197 * Mike McLagan : Routing by source
198 * Keith Owens : Do proper merging with partial SKB's in
199 * tcp_do_sendmsg to avoid burstiness.
200 * Eric Schenk : Fix fast close down bug with
201 * shutdown() followed by close().
202 * Andi Kleen : Make poll agree with SIGIO
203 * Salvatore Sanfilippo : Support SO_LINGER with linger == 1 and
204 * lingertime == 0 (RFC 793 ABORT Call)
205 * Hirokazu Takahashi : Use copy_from_user() instead of
206 * csum_and_copy_from_user() if possible.
207 *
208 * This program is free software; you can redistribute it and/or
209 * modify it under the terms of the GNU General Public License
210 * as published by the Free Software Foundation; either version
211 * 2 of the License, or(at your option) any later version.
212 *
213 * Description of States:
214 *
215 * TCP_SYN_SENT sent a connection request, waiting for ack
216 *
217 * TCP_SYN_RECV received a connection request, sent ack,
218 * waiting for final ack in three-way handshake.
219 *
220 * TCP_ESTABLISHED connection established
221 *
222 * TCP_FIN_WAIT1 our side has shutdown, waiting to complete
223 * transmission of remaining buffered data
224 *
225 * TCP_FIN_WAIT2 all buffered data sent, waiting for remote
226 * to shutdown
227 *
228 * TCP_CLOSING both sides have shutdown but we still have
229 * data we have to finish sending
230 *
231 * TCP_TIME_WAIT timeout to catch resent junk before entering
232 * closed, can only be entered from FIN_WAIT2
233 * or CLOSING. Required because the other end
234 * may not have gotten our last ACK causing it
235 * to retransmit the data packet (which we ignore)
236 *
237 * TCP_CLOSE_WAIT remote side has shutdown and is waiting for
238 * us to finish writing our data and to shutdown
239 * (we have to close() to move on to LAST_ACK)
240 *
241 * TCP_LAST_ACK out side has shutdown after remote has
242 * shutdown. There may still be data in our
243 * buffer that we have to finish sending
244 *
245 * TCP_CLOSE socket is finished
246 */
247
248#define pr_fmt(fmt) "TCP: " fmt
249
250#include <crypto/hash.h>
251#include <linux/kernel.h>
252#include <linux/module.h>
253#include <linux/types.h>
254#include <linux/fcntl.h>
255#include <linux/poll.h>
256#include <linux/inet_diag.h>
257#include <linux/init.h>
258#include <linux/fs.h>
259#include <linux/skbuff.h>
260#include <linux/scatterlist.h>
261#include <linux/splice.h>
262#include <linux/net.h>
263#include <linux/socket.h>
264#include <linux/random.h>
265#include <linux/bootmem.h>
266#include <linux/highmem.h>
267#include <linux/swap.h>
268#include <linux/cache.h>
269#include <linux/err.h>
270#include <linux/time.h>
271#include <linux/slab.h>
272
273#include <net/icmp.h>
274#include <net/inet_common.h>
275#include <net/tcp.h>
276#include <net/xfrm.h>
277#include <net/ip.h>
278#include <net/sock.h>
279
280#include <linux/uaccess.h>
281#include <asm/ioctls.h>
282#include <net/busy_poll.h>
283
284int sysctl_tcp_min_tso_segs __read_mostly = 2;
285
286int sysctl_tcp_autocorking __read_mostly = 1;
287
288struct percpu_counter tcp_orphan_count;
289EXPORT_SYMBOL_GPL(tcp_orphan_count);
290
291long sysctl_tcp_mem[3] __read_mostly;
292int sysctl_tcp_wmem[3] __read_mostly;
293int sysctl_tcp_rmem[3] __read_mostly;
294
295EXPORT_SYMBOL(sysctl_tcp_mem);
296EXPORT_SYMBOL(sysctl_tcp_rmem);
297EXPORT_SYMBOL(sysctl_tcp_wmem);
298
299atomic_long_t tcp_memory_allocated; /* Current allocated memory. */
300EXPORT_SYMBOL(tcp_memory_allocated);
301
302/*
303 * Current number of TCP sockets.
304 */
305struct percpu_counter tcp_sockets_allocated;
306EXPORT_SYMBOL(tcp_sockets_allocated);
307
308/*
309 * TCP splice context
310 */
311struct tcp_splice_state {
312 struct pipe_inode_info *pipe;
313 size_t len;
314 unsigned int flags;
315};
316
317/*
318 * Pressure flag: try to collapse.
319 * Technical note: it is used by multiple contexts non atomically.
320 * All the __sk_mem_schedule() is of this nature: accounting
321 * is strict, actions are advisory and have some latency.
322 */
323int tcp_memory_pressure __read_mostly;
324EXPORT_SYMBOL(tcp_memory_pressure);
325
326void tcp_enter_memory_pressure(struct sock *sk)
327{
328 if (!tcp_memory_pressure) {
329 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPMEMORYPRESSURES);
330 tcp_memory_pressure = 1;
331 }
332}
333EXPORT_SYMBOL(tcp_enter_memory_pressure);
334
335/* Convert seconds to retransmits based on initial and max timeout */
336static u8 secs_to_retrans(int seconds, int timeout, int rto_max)
337{
338 u8 res = 0;
339
340 if (seconds > 0) {
341 int period = timeout;
342
343 res = 1;
344 while (seconds > period && res < 255) {
345 res++;
346 timeout <<= 1;
347 if (timeout > rto_max)
348 timeout = rto_max;
349 period += timeout;
350 }
351 }
352 return res;
353}
354
355/* Convert retransmits to seconds based on initial and max timeout */
356static int retrans_to_secs(u8 retrans, int timeout, int rto_max)
357{
358 int period = 0;
359
360 if (retrans > 0) {
361 period = timeout;
362 while (--retrans) {
363 timeout <<= 1;
364 if (timeout > rto_max)
365 timeout = rto_max;
366 period += timeout;
367 }
368 }
369 return period;
370}
371
372/* Address-family independent initialization for a tcp_sock.
373 *
374 * NOTE: A lot of things set to zero explicitly by call to
375 * sk_alloc() so need not be done here.
376 */
377void tcp_init_sock(struct sock *sk)
378{
379 struct inet_connection_sock *icsk = inet_csk(sk);
380 struct tcp_sock *tp = tcp_sk(sk);
381
382 tp->out_of_order_queue = RB_ROOT;
383 tcp_init_xmit_timers(sk);
384 tcp_prequeue_init(tp);
385 INIT_LIST_HEAD(&tp->tsq_node);
386
387 icsk->icsk_rto = TCP_TIMEOUT_INIT;
388 tp->mdev_us = jiffies_to_usecs(TCP_TIMEOUT_INIT);
389 minmax_reset(&tp->rtt_min, tcp_time_stamp, ~0U);
390
391 /* So many TCP implementations out there (incorrectly) count the
392 * initial SYN frame in their delayed-ACK and congestion control
393 * algorithms that we must have the following bandaid to talk
394 * efficiently to them. -DaveM
395 */
396 tp->snd_cwnd = TCP_INIT_CWND;
397
398 /* There's a bubble in the pipe until at least the first ACK. */
399 tp->app_limited = ~0U;
400
401 /* See draft-stevens-tcpca-spec-01 for discussion of the
402 * initialization of these values.
403 */
404 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
405 tp->snd_cwnd_clamp = ~0;
406 tp->mss_cache = TCP_MSS_DEFAULT;
407
408 tp->reordering = sock_net(sk)->ipv4.sysctl_tcp_reordering;
409 tcp_enable_early_retrans(tp);
410 tcp_assign_congestion_control(sk);
411
412 tp->tsoffset = 0;
413
414 sk->sk_state = TCP_CLOSE;
415
416 sk->sk_write_space = sk_stream_write_space;
417 sock_set_flag(sk, SOCK_USE_WRITE_QUEUE);
418
419 icsk->icsk_sync_mss = tcp_sync_mss;
420
421 sk->sk_sndbuf = sysctl_tcp_wmem[1];
422 sk->sk_rcvbuf = sysctl_tcp_rmem[1];
423
424 local_bh_disable();
425 sk_sockets_allocated_inc(sk);
426 local_bh_enable();
427}
428EXPORT_SYMBOL(tcp_init_sock);
429
430static void tcp_tx_timestamp(struct sock *sk, u16 tsflags, struct sk_buff *skb)
431{
432 if (tsflags) {
433 struct skb_shared_info *shinfo = skb_shinfo(skb);
434 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
435
436 sock_tx_timestamp(sk, tsflags, &shinfo->tx_flags);
437 if (tsflags & SOF_TIMESTAMPING_TX_ACK)
438 tcb->txstamp_ack = 1;
439 if (tsflags & SOF_TIMESTAMPING_TX_RECORD_MASK)
440 shinfo->tskey = TCP_SKB_CB(skb)->seq + skb->len - 1;
441 }
442}
443
444/*
445 * Wait for a TCP event.
446 *
447 * Note that we don't need to lock the socket, as the upper poll layers
448 * take care of normal races (between the test and the event) and we don't
449 * go look at any of the socket buffers directly.
450 */
451unsigned int tcp_poll(struct file *file, struct socket *sock, poll_table *wait)
452{
453 unsigned int mask;
454 struct sock *sk = sock->sk;
455 const struct tcp_sock *tp = tcp_sk(sk);
456 int state;
457
458 sock_rps_record_flow(sk);
459
460 sock_poll_wait(file, sk_sleep(sk), wait);
461
462 state = sk_state_load(sk);
463 if (state == TCP_LISTEN)
464 return inet_csk_listen_poll(sk);
465
466 /* Socket is not locked. We are protected from async events
467 * by poll logic and correct handling of state changes
468 * made by other threads is impossible in any case.
469 */
470
471 mask = 0;
472
473 /*
474 * POLLHUP is certainly not done right. But poll() doesn't
475 * have a notion of HUP in just one direction, and for a
476 * socket the read side is more interesting.
477 *
478 * Some poll() documentation says that POLLHUP is incompatible
479 * with the POLLOUT/POLLWR flags, so somebody should check this
480 * all. But careful, it tends to be safer to return too many
481 * bits than too few, and you can easily break real applications
482 * if you don't tell them that something has hung up!
483 *
484 * Check-me.
485 *
486 * Check number 1. POLLHUP is _UNMASKABLE_ event (see UNIX98 and
487 * our fs/select.c). It means that after we received EOF,
488 * poll always returns immediately, making impossible poll() on write()
489 * in state CLOSE_WAIT. One solution is evident --- to set POLLHUP
490 * if and only if shutdown has been made in both directions.
491 * Actually, it is interesting to look how Solaris and DUX
492 * solve this dilemma. I would prefer, if POLLHUP were maskable,
493 * then we could set it on SND_SHUTDOWN. BTW examples given
494 * in Stevens' books assume exactly this behaviour, it explains
495 * why POLLHUP is incompatible with POLLOUT. --ANK
496 *
497 * NOTE. Check for TCP_CLOSE is added. The goal is to prevent
498 * blocking on fresh not-connected or disconnected socket. --ANK
499 */
500 if (sk->sk_shutdown == SHUTDOWN_MASK || state == TCP_CLOSE)
501 mask |= POLLHUP;
502 if (sk->sk_shutdown & RCV_SHUTDOWN)
503 mask |= POLLIN | POLLRDNORM | POLLRDHUP;
504
505 /* Connected or passive Fast Open socket? */
506 if (state != TCP_SYN_SENT &&
507 (state != TCP_SYN_RECV || tp->fastopen_rsk)) {
508 int target = sock_rcvlowat(sk, 0, INT_MAX);
509
510 if (tp->urg_seq == tp->copied_seq &&
511 !sock_flag(sk, SOCK_URGINLINE) &&
512 tp->urg_data)
513 target++;
514
515 if (tp->rcv_nxt - tp->copied_seq >= target)
516 mask |= POLLIN | POLLRDNORM;
517
518 if (!(sk->sk_shutdown & SEND_SHUTDOWN)) {
519 if (sk_stream_is_writeable(sk)) {
520 mask |= POLLOUT | POLLWRNORM;
521 } else { /* send SIGIO later */
522 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
523 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
524
525 /* Race breaker. If space is freed after
526 * wspace test but before the flags are set,
527 * IO signal will be lost. Memory barrier
528 * pairs with the input side.
529 */
530 smp_mb__after_atomic();
531 if (sk_stream_is_writeable(sk))
532 mask |= POLLOUT | POLLWRNORM;
533 }
534 } else
535 mask |= POLLOUT | POLLWRNORM;
536
537 if (tp->urg_data & TCP_URG_VALID)
538 mask |= POLLPRI;
539 }
540 /* This barrier is coupled with smp_wmb() in tcp_reset() */
541 smp_rmb();
542 if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
543 mask |= POLLERR;
544
545 return mask;
546}
547EXPORT_SYMBOL(tcp_poll);
548
549int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg)
550{
551 struct tcp_sock *tp = tcp_sk(sk);
552 int answ;
553 bool slow;
554
555 switch (cmd) {
556 case SIOCINQ:
557 if (sk->sk_state == TCP_LISTEN)
558 return -EINVAL;
559
560 slow = lock_sock_fast(sk);
561 answ = tcp_inq(sk);
562 unlock_sock_fast(sk, slow);
563 break;
564 case SIOCATMARK:
565 answ = tp->urg_data && tp->urg_seq == tp->copied_seq;
566 break;
567 case SIOCOUTQ:
568 if (sk->sk_state == TCP_LISTEN)
569 return -EINVAL;
570
571 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
572 answ = 0;
573 else
574 answ = tp->write_seq - tp->snd_una;
575 break;
576 case SIOCOUTQNSD:
577 if (sk->sk_state == TCP_LISTEN)
578 return -EINVAL;
579
580 if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))
581 answ = 0;
582 else
583 answ = tp->write_seq - tp->snd_nxt;
584 break;
585 default:
586 return -ENOIOCTLCMD;
587 }
588
589 return put_user(answ, (int __user *)arg);
590}
591EXPORT_SYMBOL(tcp_ioctl);
592
593static inline void tcp_mark_push(struct tcp_sock *tp, struct sk_buff *skb)
594{
595 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
596 tp->pushed_seq = tp->write_seq;
597}
598
599static inline bool forced_push(const struct tcp_sock *tp)
600{
601 return after(tp->write_seq, tp->pushed_seq + (tp->max_window >> 1));
602}
603
604static void skb_entail(struct sock *sk, struct sk_buff *skb)
605{
606 struct tcp_sock *tp = tcp_sk(sk);
607 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
608
609 skb->csum = 0;
610 tcb->seq = tcb->end_seq = tp->write_seq;
611 tcb->tcp_flags = TCPHDR_ACK;
612 tcb->sacked = 0;
613 __skb_header_release(skb);
614 tcp_add_write_queue_tail(sk, skb);
615 sk->sk_wmem_queued += skb->truesize;
616 sk_mem_charge(sk, skb->truesize);
617 if (tp->nonagle & TCP_NAGLE_PUSH)
618 tp->nonagle &= ~TCP_NAGLE_PUSH;
619
620 tcp_slow_start_after_idle_check(sk);
621}
622
623static inline void tcp_mark_urg(struct tcp_sock *tp, int flags)
624{
625 if (flags & MSG_OOB)
626 tp->snd_up = tp->write_seq;
627}
628
629/* If a not yet filled skb is pushed, do not send it if
630 * we have data packets in Qdisc or NIC queues :
631 * Because TX completion will happen shortly, it gives a chance
632 * to coalesce future sendmsg() payload into this skb, without
633 * need for a timer, and with no latency trade off.
634 * As packets containing data payload have a bigger truesize
635 * than pure acks (dataless) packets, the last checks prevent
636 * autocorking if we only have an ACK in Qdisc/NIC queues,
637 * or if TX completion was delayed after we processed ACK packet.
638 */
639static bool tcp_should_autocork(struct sock *sk, struct sk_buff *skb,
640 int size_goal)
641{
642 return skb->len < size_goal &&
643 sysctl_tcp_autocorking &&
644 skb != tcp_write_queue_head(sk) &&
645 atomic_read(&sk->sk_wmem_alloc) > skb->truesize;
646}
647
648static void tcp_push(struct sock *sk, int flags, int mss_now,
649 int nonagle, int size_goal)
650{
651 struct tcp_sock *tp = tcp_sk(sk);
652 struct sk_buff *skb;
653
654 if (!tcp_send_head(sk))
655 return;
656
657 skb = tcp_write_queue_tail(sk);
658 if (!(flags & MSG_MORE) || forced_push(tp))
659 tcp_mark_push(tp, skb);
660
661 tcp_mark_urg(tp, flags);
662
663 if (tcp_should_autocork(sk, skb, size_goal)) {
664
665 /* avoid atomic op if TSQ_THROTTLED bit is already set */
666 if (!test_bit(TSQ_THROTTLED, &sk->sk_tsq_flags)) {
667 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPAUTOCORKING);
668 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
669 }
670 /* It is possible TX completion already happened
671 * before we set TSQ_THROTTLED.
672 */
673 if (atomic_read(&sk->sk_wmem_alloc) > skb->truesize)
674 return;
675 }
676
677 if (flags & MSG_MORE)
678 nonagle = TCP_NAGLE_CORK;
679
680 __tcp_push_pending_frames(sk, mss_now, nonagle);
681}
682
683static int tcp_splice_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
684 unsigned int offset, size_t len)
685{
686 struct tcp_splice_state *tss = rd_desc->arg.data;
687 int ret;
688
689 ret = skb_splice_bits(skb, skb->sk, offset, tss->pipe,
690 min(rd_desc->count, len), tss->flags);
691 if (ret > 0)
692 rd_desc->count -= ret;
693 return ret;
694}
695
696static int __tcp_splice_read(struct sock *sk, struct tcp_splice_state *tss)
697{
698 /* Store TCP splice context information in read_descriptor_t. */
699 read_descriptor_t rd_desc = {
700 .arg.data = tss,
701 .count = tss->len,
702 };
703
704 return tcp_read_sock(sk, &rd_desc, tcp_splice_data_recv);
705}
706
707/**
708 * tcp_splice_read - splice data from TCP socket to a pipe
709 * @sock: socket to splice from
710 * @ppos: position (not valid)
711 * @pipe: pipe to splice to
712 * @len: number of bytes to splice
713 * @flags: splice modifier flags
714 *
715 * Description:
716 * Will read pages from given socket and fill them into a pipe.
717 *
718 **/
719ssize_t tcp_splice_read(struct socket *sock, loff_t *ppos,
720 struct pipe_inode_info *pipe, size_t len,
721 unsigned int flags)
722{
723 struct sock *sk = sock->sk;
724 struct tcp_splice_state tss = {
725 .pipe = pipe,
726 .len = len,
727 .flags = flags,
728 };
729 long timeo;
730 ssize_t spliced;
731 int ret;
732
733 sock_rps_record_flow(sk);
734 /*
735 * We can't seek on a socket input
736 */
737 if (unlikely(*ppos))
738 return -ESPIPE;
739
740 ret = spliced = 0;
741
742 lock_sock(sk);
743
744 timeo = sock_rcvtimeo(sk, sock->file->f_flags & O_NONBLOCK);
745 while (tss.len) {
746 ret = __tcp_splice_read(sk, &tss);
747 if (ret < 0)
748 break;
749 else if (!ret) {
750 if (spliced)
751 break;
752 if (sock_flag(sk, SOCK_DONE))
753 break;
754 if (sk->sk_err) {
755 ret = sock_error(sk);
756 break;
757 }
758 if (sk->sk_shutdown & RCV_SHUTDOWN)
759 break;
760 if (sk->sk_state == TCP_CLOSE) {
761 /*
762 * This occurs when user tries to read
763 * from never connected socket.
764 */
765 if (!sock_flag(sk, SOCK_DONE))
766 ret = -ENOTCONN;
767 break;
768 }
769 if (!timeo) {
770 ret = -EAGAIN;
771 break;
772 }
773 /* if __tcp_splice_read() got nothing while we have
774 * an skb in receive queue, we do not want to loop.
775 * This might happen with URG data.
776 */
777 if (!skb_queue_empty(&sk->sk_receive_queue))
778 break;
779 sk_wait_data(sk, &timeo, NULL);
780 if (signal_pending(current)) {
781 ret = sock_intr_errno(timeo);
782 break;
783 }
784 continue;
785 }
786 tss.len -= ret;
787 spliced += ret;
788
789 if (!timeo)
790 break;
791 release_sock(sk);
792 lock_sock(sk);
793
794 if (sk->sk_err || sk->sk_state == TCP_CLOSE ||
795 (sk->sk_shutdown & RCV_SHUTDOWN) ||
796 signal_pending(current))
797 break;
798 }
799
800 release_sock(sk);
801
802 if (spliced)
803 return spliced;
804
805 return ret;
806}
807EXPORT_SYMBOL(tcp_splice_read);
808
809struct sk_buff *sk_stream_alloc_skb(struct sock *sk, int size, gfp_t gfp,
810 bool force_schedule)
811{
812 struct sk_buff *skb;
813
814 /* The TCP header must be at least 32-bit aligned. */
815 size = ALIGN(size, 4);
816
817 if (unlikely(tcp_under_memory_pressure(sk)))
818 sk_mem_reclaim_partial(sk);
819
820 skb = alloc_skb_fclone(size + sk->sk_prot->max_header, gfp);
821 if (likely(skb)) {
822 bool mem_scheduled;
823
824 if (force_schedule) {
825 mem_scheduled = true;
826 sk_forced_mem_schedule(sk, skb->truesize);
827 } else {
828 mem_scheduled = sk_wmem_schedule(sk, skb->truesize);
829 }
830 if (likely(mem_scheduled)) {
831 skb_reserve(skb, sk->sk_prot->max_header);
832 /*
833 * Make sure that we have exactly size bytes
834 * available to the caller, no more, no less.
835 */
836 skb->reserved_tailroom = skb->end - skb->tail - size;
837 return skb;
838 }
839 __kfree_skb(skb);
840 } else {
841 sk->sk_prot->enter_memory_pressure(sk);
842 sk_stream_moderate_sndbuf(sk);
843 }
844 return NULL;
845}
846
847static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now,
848 int large_allowed)
849{
850 struct tcp_sock *tp = tcp_sk(sk);
851 u32 new_size_goal, size_goal;
852
853 if (!large_allowed || !sk_can_gso(sk))
854 return mss_now;
855
856 /* Note : tcp_tso_autosize() will eventually split this later */
857 new_size_goal = sk->sk_gso_max_size - 1 - MAX_TCP_HEADER;
858 new_size_goal = tcp_bound_to_half_wnd(tp, new_size_goal);
859
860 /* We try hard to avoid divides here */
861 size_goal = tp->gso_segs * mss_now;
862 if (unlikely(new_size_goal < size_goal ||
863 new_size_goal >= size_goal + mss_now)) {
864 tp->gso_segs = min_t(u16, new_size_goal / mss_now,
865 sk->sk_gso_max_segs);
866 size_goal = tp->gso_segs * mss_now;
867 }
868
869 return max(size_goal, mss_now);
870}
871
872static int tcp_send_mss(struct sock *sk, int *size_goal, int flags)
873{
874 int mss_now;
875
876 mss_now = tcp_current_mss(sk);
877 *size_goal = tcp_xmit_size_goal(sk, mss_now, !(flags & MSG_OOB));
878
879 return mss_now;
880}
881
882static ssize_t do_tcp_sendpages(struct sock *sk, struct page *page, int offset,
883 size_t size, int flags)
884{
885 struct tcp_sock *tp = tcp_sk(sk);
886 int mss_now, size_goal;
887 int err;
888 ssize_t copied;
889 long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
890
891 /* Wait for a connection to finish. One exception is TCP Fast Open
892 * (passive side) where data is allowed to be sent before a connection
893 * is fully established.
894 */
895 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
896 !tcp_passive_fastopen(sk)) {
897 err = sk_stream_wait_connect(sk, &timeo);
898 if (err != 0)
899 goto out_err;
900 }
901
902 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
903
904 mss_now = tcp_send_mss(sk, &size_goal, flags);
905 copied = 0;
906
907 err = -EPIPE;
908 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
909 goto out_err;
910
911 while (size > 0) {
912 struct sk_buff *skb = tcp_write_queue_tail(sk);
913 int copy, i;
914 bool can_coalesce;
915
916 if (!tcp_send_head(sk) || (copy = size_goal - skb->len) <= 0 ||
917 !tcp_skb_can_collapse_to(skb)) {
918new_segment:
919 if (!sk_stream_memory_free(sk))
920 goto wait_for_sndbuf;
921
922 skb = sk_stream_alloc_skb(sk, 0, sk->sk_allocation,
923 skb_queue_empty(&sk->sk_write_queue));
924 if (!skb)
925 goto wait_for_memory;
926
927 skb_entail(sk, skb);
928 copy = size_goal;
929 }
930
931 if (copy > size)
932 copy = size;
933
934 i = skb_shinfo(skb)->nr_frags;
935 can_coalesce = skb_can_coalesce(skb, i, page, offset);
936 if (!can_coalesce && i >= sysctl_max_skb_frags) {
937 tcp_mark_push(tp, skb);
938 goto new_segment;
939 }
940 if (!sk_wmem_schedule(sk, copy))
941 goto wait_for_memory;
942
943 if (can_coalesce) {
944 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
945 } else {
946 get_page(page);
947 skb_fill_page_desc(skb, i, page, offset, copy);
948 }
949 skb_shinfo(skb)->tx_flags |= SKBTX_SHARED_FRAG;
950
951 skb->len += copy;
952 skb->data_len += copy;
953 skb->truesize += copy;
954 sk->sk_wmem_queued += copy;
955 sk_mem_charge(sk, copy);
956 skb->ip_summed = CHECKSUM_PARTIAL;
957 tp->write_seq += copy;
958 TCP_SKB_CB(skb)->end_seq += copy;
959 tcp_skb_pcount_set(skb, 0);
960
961 if (!copied)
962 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
963
964 copied += copy;
965 offset += copy;
966 size -= copy;
967 if (!size) {
968 tcp_tx_timestamp(sk, sk->sk_tsflags, skb);
969 goto out;
970 }
971
972 if (skb->len < size_goal || (flags & MSG_OOB))
973 continue;
974
975 if (forced_push(tp)) {
976 tcp_mark_push(tp, skb);
977 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
978 } else if (skb == tcp_send_head(sk))
979 tcp_push_one(sk, mss_now);
980 continue;
981
982wait_for_sndbuf:
983 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
984wait_for_memory:
985 tcp_push(sk, flags & ~MSG_MORE, mss_now,
986 TCP_NAGLE_PUSH, size_goal);
987
988 err = sk_stream_wait_memory(sk, &timeo);
989 if (err != 0)
990 goto do_error;
991
992 mss_now = tcp_send_mss(sk, &size_goal, flags);
993 }
994
995out:
996 if (copied && !(flags & MSG_SENDPAGE_NOTLAST))
997 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
998 return copied;
999
1000do_error:
1001 if (copied)
1002 goto out;
1003out_err:
1004 /* make sure we wake any epoll edge trigger waiter */
1005 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1006 err == -EAGAIN)) {
1007 sk->sk_write_space(sk);
1008 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1009 }
1010 return sk_stream_error(sk, flags, err);
1011}
1012
1013int tcp_sendpage(struct sock *sk, struct page *page, int offset,
1014 size_t size, int flags)
1015{
1016 ssize_t res;
1017
1018 if (!(sk->sk_route_caps & NETIF_F_SG) ||
1019 !sk_check_csum_caps(sk))
1020 return sock_no_sendpage(sk->sk_socket, page, offset, size,
1021 flags);
1022
1023 lock_sock(sk);
1024
1025 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1026
1027 res = do_tcp_sendpages(sk, page, offset, size, flags);
1028 release_sock(sk);
1029 return res;
1030}
1031EXPORT_SYMBOL(tcp_sendpage);
1032
1033/* Do not bother using a page frag for very small frames.
1034 * But use this heuristic only for the first skb in write queue.
1035 *
1036 * Having no payload in skb->head allows better SACK shifting
1037 * in tcp_shift_skb_data(), reducing sack/rack overhead, because
1038 * write queue has less skbs.
1039 * Each skb can hold up to MAX_SKB_FRAGS * 32Kbytes, or ~0.5 MB.
1040 * This also speeds up tso_fragment(), since it wont fallback
1041 * to tcp_fragment().
1042 */
1043static int linear_payload_sz(bool first_skb)
1044{
1045 if (first_skb)
1046 return SKB_WITH_OVERHEAD(2048 - MAX_TCP_HEADER);
1047 return 0;
1048}
1049
1050static int select_size(const struct sock *sk, bool sg, bool first_skb)
1051{
1052 const struct tcp_sock *tp = tcp_sk(sk);
1053 int tmp = tp->mss_cache;
1054
1055 if (sg) {
1056 if (sk_can_gso(sk)) {
1057 tmp = linear_payload_sz(first_skb);
1058 } else {
1059 int pgbreak = SKB_MAX_HEAD(MAX_TCP_HEADER);
1060
1061 if (tmp >= pgbreak &&
1062 tmp <= pgbreak + (MAX_SKB_FRAGS - 1) * PAGE_SIZE)
1063 tmp = pgbreak;
1064 }
1065 }
1066
1067 return tmp;
1068}
1069
1070void tcp_free_fastopen_req(struct tcp_sock *tp)
1071{
1072 if (tp->fastopen_req) {
1073 kfree(tp->fastopen_req);
1074 tp->fastopen_req = NULL;
1075 }
1076}
1077
1078static int tcp_sendmsg_fastopen(struct sock *sk, struct msghdr *msg,
1079 int *copied, size_t size)
1080{
1081 struct tcp_sock *tp = tcp_sk(sk);
1082 int err, flags;
1083
1084 if (!(sysctl_tcp_fastopen & TFO_CLIENT_ENABLE))
1085 return -EOPNOTSUPP;
1086 if (tp->fastopen_req)
1087 return -EALREADY; /* Another Fast Open is in progress */
1088
1089 tp->fastopen_req = kzalloc(sizeof(struct tcp_fastopen_request),
1090 sk->sk_allocation);
1091 if (unlikely(!tp->fastopen_req))
1092 return -ENOBUFS;
1093 tp->fastopen_req->data = msg;
1094 tp->fastopen_req->size = size;
1095
1096 flags = (msg->msg_flags & MSG_DONTWAIT) ? O_NONBLOCK : 0;
1097 err = __inet_stream_connect(sk->sk_socket, msg->msg_name,
1098 msg->msg_namelen, flags);
1099 *copied = tp->fastopen_req->copied;
1100 tcp_free_fastopen_req(tp);
1101 return err;
1102}
1103
1104int tcp_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1105{
1106 struct tcp_sock *tp = tcp_sk(sk);
1107 struct sk_buff *skb;
1108 struct sockcm_cookie sockc;
1109 int flags, err, copied = 0;
1110 int mss_now = 0, size_goal, copied_syn = 0;
1111 bool process_backlog = false;
1112 bool sg;
1113 long timeo;
1114
1115 lock_sock(sk);
1116
1117 flags = msg->msg_flags;
1118 if (flags & MSG_FASTOPEN) {
1119 err = tcp_sendmsg_fastopen(sk, msg, &copied_syn, size);
1120 if (err == -EINPROGRESS && copied_syn > 0)
1121 goto out;
1122 else if (err)
1123 goto out_err;
1124 }
1125
1126 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1127
1128 tcp_rate_check_app_limited(sk); /* is sending application-limited? */
1129
1130 /* Wait for a connection to finish. One exception is TCP Fast Open
1131 * (passive side) where data is allowed to be sent before a connection
1132 * is fully established.
1133 */
1134 if (((1 << sk->sk_state) & ~(TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) &&
1135 !tcp_passive_fastopen(sk)) {
1136 err = sk_stream_wait_connect(sk, &timeo);
1137 if (err != 0)
1138 goto do_error;
1139 }
1140
1141 if (unlikely(tp->repair)) {
1142 if (tp->repair_queue == TCP_RECV_QUEUE) {
1143 copied = tcp_send_rcvq(sk, msg, size);
1144 goto out_nopush;
1145 }
1146
1147 err = -EINVAL;
1148 if (tp->repair_queue == TCP_NO_QUEUE)
1149 goto out_err;
1150
1151 /* 'common' sending to sendq */
1152 }
1153
1154 sockc.tsflags = sk->sk_tsflags;
1155 if (msg->msg_controllen) {
1156 err = sock_cmsg_send(sk, msg, &sockc);
1157 if (unlikely(err)) {
1158 err = -EINVAL;
1159 goto out_err;
1160 }
1161 }
1162
1163 /* This should be in poll */
1164 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1165
1166 /* Ok commence sending. */
1167 copied = 0;
1168
1169restart:
1170 mss_now = tcp_send_mss(sk, &size_goal, flags);
1171
1172 err = -EPIPE;
1173 if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN))
1174 goto do_error;
1175
1176 sg = !!(sk->sk_route_caps & NETIF_F_SG);
1177
1178 while (msg_data_left(msg)) {
1179 int copy = 0;
1180 int max = size_goal;
1181
1182 skb = tcp_write_queue_tail(sk);
1183 if (tcp_send_head(sk)) {
1184 if (skb->ip_summed == CHECKSUM_NONE)
1185 max = mss_now;
1186 copy = max - skb->len;
1187 }
1188
1189 if (copy <= 0 || !tcp_skb_can_collapse_to(skb)) {
1190 bool first_skb;
1191
1192new_segment:
1193 /* Allocate new segment. If the interface is SG,
1194 * allocate skb fitting to single page.
1195 */
1196 if (!sk_stream_memory_free(sk))
1197 goto wait_for_sndbuf;
1198
1199 if (process_backlog && sk_flush_backlog(sk)) {
1200 process_backlog = false;
1201 goto restart;
1202 }
1203 first_skb = skb_queue_empty(&sk->sk_write_queue);
1204 skb = sk_stream_alloc_skb(sk,
1205 select_size(sk, sg, first_skb),
1206 sk->sk_allocation,
1207 first_skb);
1208 if (!skb)
1209 goto wait_for_memory;
1210
1211 process_backlog = true;
1212 /*
1213 * Check whether we can use HW checksum.
1214 */
1215 if (sk_check_csum_caps(sk))
1216 skb->ip_summed = CHECKSUM_PARTIAL;
1217
1218 skb_entail(sk, skb);
1219 copy = size_goal;
1220 max = size_goal;
1221
1222 /* All packets are restored as if they have
1223 * already been sent. skb_mstamp isn't set to
1224 * avoid wrong rtt estimation.
1225 */
1226 if (tp->repair)
1227 TCP_SKB_CB(skb)->sacked |= TCPCB_REPAIRED;
1228 }
1229
1230 /* Try to append data to the end of skb. */
1231 if (copy > msg_data_left(msg))
1232 copy = msg_data_left(msg);
1233
1234 /* Where to copy to? */
1235 if (skb_availroom(skb) > 0) {
1236 /* We have some space in skb head. Superb! */
1237 copy = min_t(int, copy, skb_availroom(skb));
1238 err = skb_add_data_nocache(sk, skb, &msg->msg_iter, copy);
1239 if (err)
1240 goto do_fault;
1241 } else {
1242 bool merge = true;
1243 int i = skb_shinfo(skb)->nr_frags;
1244 struct page_frag *pfrag = sk_page_frag(sk);
1245
1246 if (!sk_page_frag_refill(sk, pfrag))
1247 goto wait_for_memory;
1248
1249 if (!skb_can_coalesce(skb, i, pfrag->page,
1250 pfrag->offset)) {
1251 if (i >= sysctl_max_skb_frags || !sg) {
1252 tcp_mark_push(tp, skb);
1253 goto new_segment;
1254 }
1255 merge = false;
1256 }
1257
1258 copy = min_t(int, copy, pfrag->size - pfrag->offset);
1259
1260 if (!sk_wmem_schedule(sk, copy))
1261 goto wait_for_memory;
1262
1263 err = skb_copy_to_page_nocache(sk, &msg->msg_iter, skb,
1264 pfrag->page,
1265 pfrag->offset,
1266 copy);
1267 if (err)
1268 goto do_error;
1269
1270 /* Update the skb. */
1271 if (merge) {
1272 skb_frag_size_add(&skb_shinfo(skb)->frags[i - 1], copy);
1273 } else {
1274 skb_fill_page_desc(skb, i, pfrag->page,
1275 pfrag->offset, copy);
1276 get_page(pfrag->page);
1277 }
1278 pfrag->offset += copy;
1279 }
1280
1281 if (!copied)
1282 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_PSH;
1283
1284 tp->write_seq += copy;
1285 TCP_SKB_CB(skb)->end_seq += copy;
1286 tcp_skb_pcount_set(skb, 0);
1287
1288 copied += copy;
1289 if (!msg_data_left(msg)) {
1290 tcp_tx_timestamp(sk, sockc.tsflags, skb);
1291 if (unlikely(flags & MSG_EOR))
1292 TCP_SKB_CB(skb)->eor = 1;
1293 goto out;
1294 }
1295
1296 if (skb->len < max || (flags & MSG_OOB) || unlikely(tp->repair))
1297 continue;
1298
1299 if (forced_push(tp)) {
1300 tcp_mark_push(tp, skb);
1301 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_PUSH);
1302 } else if (skb == tcp_send_head(sk))
1303 tcp_push_one(sk, mss_now);
1304 continue;
1305
1306wait_for_sndbuf:
1307 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1308wait_for_memory:
1309 if (copied)
1310 tcp_push(sk, flags & ~MSG_MORE, mss_now,
1311 TCP_NAGLE_PUSH, size_goal);
1312
1313 err = sk_stream_wait_memory(sk, &timeo);
1314 if (err != 0)
1315 goto do_error;
1316
1317 mss_now = tcp_send_mss(sk, &size_goal, flags);
1318 }
1319
1320out:
1321 if (copied)
1322 tcp_push(sk, flags, mss_now, tp->nonagle, size_goal);
1323out_nopush:
1324 release_sock(sk);
1325 return copied + copied_syn;
1326
1327do_fault:
1328 if (!skb->len) {
1329 tcp_unlink_write_queue(skb, sk);
1330 /* It is the one place in all of TCP, except connection
1331 * reset, where we can be unlinking the send_head.
1332 */
1333 tcp_check_send_head(sk, skb);
1334 sk_wmem_free_skb(sk, skb);
1335 }
1336
1337do_error:
1338 if (copied + copied_syn)
1339 goto out;
1340out_err:
1341 err = sk_stream_error(sk, flags, err);
1342 /* make sure we wake any epoll edge trigger waiter */
1343 if (unlikely(skb_queue_len(&sk->sk_write_queue) == 0 &&
1344 err == -EAGAIN)) {
1345 sk->sk_write_space(sk);
1346 tcp_chrono_stop(sk, TCP_CHRONO_SNDBUF_LIMITED);
1347 }
1348 release_sock(sk);
1349 return err;
1350}
1351EXPORT_SYMBOL(tcp_sendmsg);
1352
1353/*
1354 * Handle reading urgent data. BSD has very simple semantics for
1355 * this, no blocking and very strange errors 8)
1356 */
1357
1358static int tcp_recv_urg(struct sock *sk, struct msghdr *msg, int len, int flags)
1359{
1360 struct tcp_sock *tp = tcp_sk(sk);
1361
1362 /* No URG data to read. */
1363 if (sock_flag(sk, SOCK_URGINLINE) || !tp->urg_data ||
1364 tp->urg_data == TCP_URG_READ)
1365 return -EINVAL; /* Yes this is right ! */
1366
1367 if (sk->sk_state == TCP_CLOSE && !sock_flag(sk, SOCK_DONE))
1368 return -ENOTCONN;
1369
1370 if (tp->urg_data & TCP_URG_VALID) {
1371 int err = 0;
1372 char c = tp->urg_data;
1373
1374 if (!(flags & MSG_PEEK))
1375 tp->urg_data = TCP_URG_READ;
1376
1377 /* Read urgent data. */
1378 msg->msg_flags |= MSG_OOB;
1379
1380 if (len > 0) {
1381 if (!(flags & MSG_TRUNC))
1382 err = memcpy_to_msg(msg, &c, 1);
1383 len = 1;
1384 } else
1385 msg->msg_flags |= MSG_TRUNC;
1386
1387 return err ? -EFAULT : len;
1388 }
1389
1390 if (sk->sk_state == TCP_CLOSE || (sk->sk_shutdown & RCV_SHUTDOWN))
1391 return 0;
1392
1393 /* Fixed the recv(..., MSG_OOB) behaviour. BSD docs and
1394 * the available implementations agree in this case:
1395 * this call should never block, independent of the
1396 * blocking state of the socket.
1397 * Mike <pall@rz.uni-karlsruhe.de>
1398 */
1399 return -EAGAIN;
1400}
1401
1402static int tcp_peek_sndq(struct sock *sk, struct msghdr *msg, int len)
1403{
1404 struct sk_buff *skb;
1405 int copied = 0, err = 0;
1406
1407 /* XXX -- need to support SO_PEEK_OFF */
1408
1409 skb_queue_walk(&sk->sk_write_queue, skb) {
1410 err = skb_copy_datagram_msg(skb, 0, msg, skb->len);
1411 if (err)
1412 break;
1413
1414 copied += skb->len;
1415 }
1416
1417 return err ?: copied;
1418}
1419
1420/* Clean up the receive buffer for full frames taken by the user,
1421 * then send an ACK if necessary. COPIED is the number of bytes
1422 * tcp_recvmsg has given to the user so far, it speeds up the
1423 * calculation of whether or not we must ACK for the sake of
1424 * a window update.
1425 */
1426static void tcp_cleanup_rbuf(struct sock *sk, int copied)
1427{
1428 struct tcp_sock *tp = tcp_sk(sk);
1429 bool time_to_ack = false;
1430
1431 struct sk_buff *skb = skb_peek(&sk->sk_receive_queue);
1432
1433 WARN(skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq),
1434 "cleanup rbuf bug: copied %X seq %X rcvnxt %X\n",
1435 tp->copied_seq, TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt);
1436
1437 if (inet_csk_ack_scheduled(sk)) {
1438 const struct inet_connection_sock *icsk = inet_csk(sk);
1439 /* Delayed ACKs frequently hit locked sockets during bulk
1440 * receive. */
1441 if (icsk->icsk_ack.blocked ||
1442 /* Once-per-two-segments ACK was not sent by tcp_input.c */
1443 tp->rcv_nxt - tp->rcv_wup > icsk->icsk_ack.rcv_mss ||
1444 /*
1445 * If this read emptied read buffer, we send ACK, if
1446 * connection is not bidirectional, user drained
1447 * receive buffer and there was a small segment
1448 * in queue.
1449 */
1450 (copied > 0 &&
1451 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED2) ||
1452 ((icsk->icsk_ack.pending & ICSK_ACK_PUSHED) &&
1453 !icsk->icsk_ack.pingpong)) &&
1454 !atomic_read(&sk->sk_rmem_alloc)))
1455 time_to_ack = true;
1456 }
1457
1458 /* We send an ACK if we can now advertise a non-zero window
1459 * which has been raised "significantly".
1460 *
1461 * Even if window raised up to infinity, do not send window open ACK
1462 * in states, where we will not receive more. It is useless.
1463 */
1464 if (copied > 0 && !time_to_ack && !(sk->sk_shutdown & RCV_SHUTDOWN)) {
1465 __u32 rcv_window_now = tcp_receive_window(tp);
1466
1467 /* Optimize, __tcp_select_window() is not cheap. */
1468 if (2*rcv_window_now <= tp->window_clamp) {
1469 __u32 new_window = __tcp_select_window(sk);
1470
1471 /* Send ACK now, if this read freed lots of space
1472 * in our buffer. Certainly, new_window is new window.
1473 * We can advertise it now, if it is not less than current one.
1474 * "Lots" means "at least twice" here.
1475 */
1476 if (new_window && new_window >= 2 * rcv_window_now)
1477 time_to_ack = true;
1478 }
1479 }
1480 if (time_to_ack)
1481 tcp_send_ack(sk);
1482}
1483
1484static void tcp_prequeue_process(struct sock *sk)
1485{
1486 struct sk_buff *skb;
1487 struct tcp_sock *tp = tcp_sk(sk);
1488
1489 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPREQUEUED);
1490
1491 while ((skb = __skb_dequeue(&tp->ucopy.prequeue)) != NULL)
1492 sk_backlog_rcv(sk, skb);
1493
1494 /* Clear memory counter. */
1495 tp->ucopy.memory = 0;
1496}
1497
1498static struct sk_buff *tcp_recv_skb(struct sock *sk, u32 seq, u32 *off)
1499{
1500 struct sk_buff *skb;
1501 u32 offset;
1502
1503 while ((skb = skb_peek(&sk->sk_receive_queue)) != NULL) {
1504 offset = seq - TCP_SKB_CB(skb)->seq;
1505 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1506 pr_err_once("%s: found a SYN, please report !\n", __func__);
1507 offset--;
1508 }
1509 if (offset < skb->len || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) {
1510 *off = offset;
1511 return skb;
1512 }
1513 /* This looks weird, but this can happen if TCP collapsing
1514 * splitted a fat GRO packet, while we released socket lock
1515 * in skb_splice_bits()
1516 */
1517 sk_eat_skb(sk, skb);
1518 }
1519 return NULL;
1520}
1521
1522/*
1523 * This routine provides an alternative to tcp_recvmsg() for routines
1524 * that would like to handle copying from skbuffs directly in 'sendfile'
1525 * fashion.
1526 * Note:
1527 * - It is assumed that the socket was locked by the caller.
1528 * - The routine does not block.
1529 * - At present, there is no support for reading OOB data
1530 * or for 'peeking' the socket using this routine
1531 * (although both would be easy to implement).
1532 */
1533int tcp_read_sock(struct sock *sk, read_descriptor_t *desc,
1534 sk_read_actor_t recv_actor)
1535{
1536 struct sk_buff *skb;
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 u32 seq = tp->copied_seq;
1539 u32 offset;
1540 int copied = 0;
1541
1542 if (sk->sk_state == TCP_LISTEN)
1543 return -ENOTCONN;
1544 while ((skb = tcp_recv_skb(sk, seq, &offset)) != NULL) {
1545 if (offset < skb->len) {
1546 int used;
1547 size_t len;
1548
1549 len = skb->len - offset;
1550 /* Stop reading if we hit a patch of urgent data */
1551 if (tp->urg_data) {
1552 u32 urg_offset = tp->urg_seq - seq;
1553 if (urg_offset < len)
1554 len = urg_offset;
1555 if (!len)
1556 break;
1557 }
1558 used = recv_actor(desc, skb, offset, len);
1559 if (used <= 0) {
1560 if (!copied)
1561 copied = used;
1562 break;
1563 } else if (used <= len) {
1564 seq += used;
1565 copied += used;
1566 offset += used;
1567 }
1568 /* If recv_actor drops the lock (e.g. TCP splice
1569 * receive) the skb pointer might be invalid when
1570 * getting here: tcp_collapse might have deleted it
1571 * while aggregating skbs from the socket queue.
1572 */
1573 skb = tcp_recv_skb(sk, seq - 1, &offset);
1574 if (!skb)
1575 break;
1576 /* TCP coalescing might have appended data to the skb.
1577 * Try to splice more frags
1578 */
1579 if (offset + 1 != skb->len)
1580 continue;
1581 }
1582 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) {
1583 sk_eat_skb(sk, skb);
1584 ++seq;
1585 break;
1586 }
1587 sk_eat_skb(sk, skb);
1588 if (!desc->count)
1589 break;
1590 tp->copied_seq = seq;
1591 }
1592 tp->copied_seq = seq;
1593
1594 tcp_rcv_space_adjust(sk);
1595
1596 /* Clean up data we have read: This will do ACK frames. */
1597 if (copied > 0) {
1598 tcp_recv_skb(sk, seq, &offset);
1599 tcp_cleanup_rbuf(sk, copied);
1600 }
1601 return copied;
1602}
1603EXPORT_SYMBOL(tcp_read_sock);
1604
1605int tcp_peek_len(struct socket *sock)
1606{
1607 return tcp_inq(sock->sk);
1608}
1609EXPORT_SYMBOL(tcp_peek_len);
1610
1611/*
1612 * This routine copies from a sock struct into the user buffer.
1613 *
1614 * Technical note: in 2.3 we work on _locked_ socket, so that
1615 * tricks with *seq access order and skb->users are not required.
1616 * Probably, code can be easily improved even more.
1617 */
1618
1619int tcp_recvmsg(struct sock *sk, struct msghdr *msg, size_t len, int nonblock,
1620 int flags, int *addr_len)
1621{
1622 struct tcp_sock *tp = tcp_sk(sk);
1623 int copied = 0;
1624 u32 peek_seq;
1625 u32 *seq;
1626 unsigned long used;
1627 int err;
1628 int target; /* Read at least this many bytes */
1629 long timeo;
1630 struct task_struct *user_recv = NULL;
1631 struct sk_buff *skb, *last;
1632 u32 urg_hole = 0;
1633
1634 if (unlikely(flags & MSG_ERRQUEUE))
1635 return inet_recv_error(sk, msg, len, addr_len);
1636
1637 if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue) &&
1638 (sk->sk_state == TCP_ESTABLISHED))
1639 sk_busy_loop(sk, nonblock);
1640
1641 lock_sock(sk);
1642
1643 err = -ENOTCONN;
1644 if (sk->sk_state == TCP_LISTEN)
1645 goto out;
1646
1647 timeo = sock_rcvtimeo(sk, nonblock);
1648
1649 /* Urgent data needs to be handled specially. */
1650 if (flags & MSG_OOB)
1651 goto recv_urg;
1652
1653 if (unlikely(tp->repair)) {
1654 err = -EPERM;
1655 if (!(flags & MSG_PEEK))
1656 goto out;
1657
1658 if (tp->repair_queue == TCP_SEND_QUEUE)
1659 goto recv_sndq;
1660
1661 err = -EINVAL;
1662 if (tp->repair_queue == TCP_NO_QUEUE)
1663 goto out;
1664
1665 /* 'common' recv queue MSG_PEEK-ing */
1666 }
1667
1668 seq = &tp->copied_seq;
1669 if (flags & MSG_PEEK) {
1670 peek_seq = tp->copied_seq;
1671 seq = &peek_seq;
1672 }
1673
1674 target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1675
1676 do {
1677 u32 offset;
1678
1679 /* Are we at urgent data? Stop if we have read anything or have SIGURG pending. */
1680 if (tp->urg_data && tp->urg_seq == *seq) {
1681 if (copied)
1682 break;
1683 if (signal_pending(current)) {
1684 copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
1685 break;
1686 }
1687 }
1688
1689 /* Next get a buffer. */
1690
1691 last = skb_peek_tail(&sk->sk_receive_queue);
1692 skb_queue_walk(&sk->sk_receive_queue, skb) {
1693 last = skb;
1694 /* Now that we have two receive queues this
1695 * shouldn't happen.
1696 */
1697 if (WARN(before(*seq, TCP_SKB_CB(skb)->seq),
1698 "recvmsg bug: copied %X seq %X rcvnxt %X fl %X\n",
1699 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt,
1700 flags))
1701 break;
1702
1703 offset = *seq - TCP_SKB_CB(skb)->seq;
1704 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
1705 pr_err_once("%s: found a SYN, please report !\n", __func__);
1706 offset--;
1707 }
1708 if (offset < skb->len)
1709 goto found_ok_skb;
1710 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1711 goto found_fin_ok;
1712 WARN(!(flags & MSG_PEEK),
1713 "recvmsg bug 2: copied %X seq %X rcvnxt %X fl %X\n",
1714 *seq, TCP_SKB_CB(skb)->seq, tp->rcv_nxt, flags);
1715 }
1716
1717 /* Well, if we have backlog, try to process it now yet. */
1718
1719 if (copied >= target && !sk->sk_backlog.tail)
1720 break;
1721
1722 if (copied) {
1723 if (sk->sk_err ||
1724 sk->sk_state == TCP_CLOSE ||
1725 (sk->sk_shutdown & RCV_SHUTDOWN) ||
1726 !timeo ||
1727 signal_pending(current))
1728 break;
1729 } else {
1730 if (sock_flag(sk, SOCK_DONE))
1731 break;
1732
1733 if (sk->sk_err) {
1734 copied = sock_error(sk);
1735 break;
1736 }
1737
1738 if (sk->sk_shutdown & RCV_SHUTDOWN)
1739 break;
1740
1741 if (sk->sk_state == TCP_CLOSE) {
1742 if (!sock_flag(sk, SOCK_DONE)) {
1743 /* This occurs when user tries to read
1744 * from never connected socket.
1745 */
1746 copied = -ENOTCONN;
1747 break;
1748 }
1749 break;
1750 }
1751
1752 if (!timeo) {
1753 copied = -EAGAIN;
1754 break;
1755 }
1756
1757 if (signal_pending(current)) {
1758 copied = sock_intr_errno(timeo);
1759 break;
1760 }
1761 }
1762
1763 tcp_cleanup_rbuf(sk, copied);
1764
1765 if (!sysctl_tcp_low_latency && tp->ucopy.task == user_recv) {
1766 /* Install new reader */
1767 if (!user_recv && !(flags & (MSG_TRUNC | MSG_PEEK))) {
1768 user_recv = current;
1769 tp->ucopy.task = user_recv;
1770 tp->ucopy.msg = msg;
1771 }
1772
1773 tp->ucopy.len = len;
1774
1775 WARN_ON(tp->copied_seq != tp->rcv_nxt &&
1776 !(flags & (MSG_PEEK | MSG_TRUNC)));
1777
1778 /* Ugly... If prequeue is not empty, we have to
1779 * process it before releasing socket, otherwise
1780 * order will be broken at second iteration.
1781 * More elegant solution is required!!!
1782 *
1783 * Look: we have the following (pseudo)queues:
1784 *
1785 * 1. packets in flight
1786 * 2. backlog
1787 * 3. prequeue
1788 * 4. receive_queue
1789 *
1790 * Each queue can be processed only if the next ones
1791 * are empty. At this point we have empty receive_queue.
1792 * But prequeue _can_ be not empty after 2nd iteration,
1793 * when we jumped to start of loop because backlog
1794 * processing added something to receive_queue.
1795 * We cannot release_sock(), because backlog contains
1796 * packets arrived _after_ prequeued ones.
1797 *
1798 * Shortly, algorithm is clear --- to process all
1799 * the queues in order. We could make it more directly,
1800 * requeueing packets from backlog to prequeue, if
1801 * is not empty. It is more elegant, but eats cycles,
1802 * unfortunately.
1803 */
1804 if (!skb_queue_empty(&tp->ucopy.prequeue))
1805 goto do_prequeue;
1806
1807 /* __ Set realtime policy in scheduler __ */
1808 }
1809
1810 if (copied >= target) {
1811 /* Do not sleep, just process backlog. */
1812 release_sock(sk);
1813 lock_sock(sk);
1814 } else {
1815 sk_wait_data(sk, &timeo, last);
1816 }
1817
1818 if (user_recv) {
1819 int chunk;
1820
1821 /* __ Restore normal policy in scheduler __ */
1822
1823 chunk = len - tp->ucopy.len;
1824 if (chunk != 0) {
1825 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMBACKLOG, chunk);
1826 len -= chunk;
1827 copied += chunk;
1828 }
1829
1830 if (tp->rcv_nxt == tp->copied_seq &&
1831 !skb_queue_empty(&tp->ucopy.prequeue)) {
1832do_prequeue:
1833 tcp_prequeue_process(sk);
1834
1835 chunk = len - tp->ucopy.len;
1836 if (chunk != 0) {
1837 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1838 len -= chunk;
1839 copied += chunk;
1840 }
1841 }
1842 }
1843 if ((flags & MSG_PEEK) &&
1844 (peek_seq - copied - urg_hole != tp->copied_seq)) {
1845 net_dbg_ratelimited("TCP(%s:%d): Application bug, race in MSG_PEEK\n",
1846 current->comm,
1847 task_pid_nr(current));
1848 peek_seq = tp->copied_seq;
1849 }
1850 continue;
1851
1852 found_ok_skb:
1853 /* Ok so how much can we use? */
1854 used = skb->len - offset;
1855 if (len < used)
1856 used = len;
1857
1858 /* Do we have urgent data here? */
1859 if (tp->urg_data) {
1860 u32 urg_offset = tp->urg_seq - *seq;
1861 if (urg_offset < used) {
1862 if (!urg_offset) {
1863 if (!sock_flag(sk, SOCK_URGINLINE)) {
1864 ++*seq;
1865 urg_hole++;
1866 offset++;
1867 used--;
1868 if (!used)
1869 goto skip_copy;
1870 }
1871 } else
1872 used = urg_offset;
1873 }
1874 }
1875
1876 if (!(flags & MSG_TRUNC)) {
1877 err = skb_copy_datagram_msg(skb, offset, msg, used);
1878 if (err) {
1879 /* Exception. Bailout! */
1880 if (!copied)
1881 copied = -EFAULT;
1882 break;
1883 }
1884 }
1885
1886 *seq += used;
1887 copied += used;
1888 len -= used;
1889
1890 tcp_rcv_space_adjust(sk);
1891
1892skip_copy:
1893 if (tp->urg_data && after(tp->copied_seq, tp->urg_seq)) {
1894 tp->urg_data = 0;
1895 tcp_fast_path_check(sk);
1896 }
1897 if (used + offset < skb->len)
1898 continue;
1899
1900 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1901 goto found_fin_ok;
1902 if (!(flags & MSG_PEEK))
1903 sk_eat_skb(sk, skb);
1904 continue;
1905
1906 found_fin_ok:
1907 /* Process the FIN. */
1908 ++*seq;
1909 if (!(flags & MSG_PEEK))
1910 sk_eat_skb(sk, skb);
1911 break;
1912 } while (len > 0);
1913
1914 if (user_recv) {
1915 if (!skb_queue_empty(&tp->ucopy.prequeue)) {
1916 int chunk;
1917
1918 tp->ucopy.len = copied > 0 ? len : 0;
1919
1920 tcp_prequeue_process(sk);
1921
1922 if (copied > 0 && (chunk = len - tp->ucopy.len) != 0) {
1923 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPDIRECTCOPYFROMPREQUEUE, chunk);
1924 len -= chunk;
1925 copied += chunk;
1926 }
1927 }
1928
1929 tp->ucopy.task = NULL;
1930 tp->ucopy.len = 0;
1931 }
1932
1933 /* According to UNIX98, msg_name/msg_namelen are ignored
1934 * on connected socket. I was just happy when found this 8) --ANK
1935 */
1936
1937 /* Clean up data we have read: This will do ACK frames. */
1938 tcp_cleanup_rbuf(sk, copied);
1939
1940 release_sock(sk);
1941 return copied;
1942
1943out:
1944 release_sock(sk);
1945 return err;
1946
1947recv_urg:
1948 err = tcp_recv_urg(sk, msg, len, flags);
1949 goto out;
1950
1951recv_sndq:
1952 err = tcp_peek_sndq(sk, msg, len);
1953 goto out;
1954}
1955EXPORT_SYMBOL(tcp_recvmsg);
1956
1957void tcp_set_state(struct sock *sk, int state)
1958{
1959 int oldstate = sk->sk_state;
1960
1961 switch (state) {
1962 case TCP_ESTABLISHED:
1963 if (oldstate != TCP_ESTABLISHED)
1964 TCP_INC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1965 break;
1966
1967 case TCP_CLOSE:
1968 if (oldstate == TCP_CLOSE_WAIT || oldstate == TCP_ESTABLISHED)
1969 TCP_INC_STATS(sock_net(sk), TCP_MIB_ESTABRESETS);
1970
1971 sk->sk_prot->unhash(sk);
1972 if (inet_csk(sk)->icsk_bind_hash &&
1973 !(sk->sk_userlocks & SOCK_BINDPORT_LOCK))
1974 inet_put_port(sk);
1975 /* fall through */
1976 default:
1977 if (oldstate == TCP_ESTABLISHED)
1978 TCP_DEC_STATS(sock_net(sk), TCP_MIB_CURRESTAB);
1979 }
1980
1981 /* Change state AFTER socket is unhashed to avoid closed
1982 * socket sitting in hash tables.
1983 */
1984 sk_state_store(sk, state);
1985
1986#ifdef STATE_TRACE
1987 SOCK_DEBUG(sk, "TCP sk=%p, State %s -> %s\n", sk, statename[oldstate], statename[state]);
1988#endif
1989}
1990EXPORT_SYMBOL_GPL(tcp_set_state);
1991
1992/*
1993 * State processing on a close. This implements the state shift for
1994 * sending our FIN frame. Note that we only send a FIN for some
1995 * states. A shutdown() may have already sent the FIN, or we may be
1996 * closed.
1997 */
1998
1999static const unsigned char new_state[16] = {
2000 /* current state: new state: action: */
2001 [0 /* (Invalid) */] = TCP_CLOSE,
2002 [TCP_ESTABLISHED] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2003 [TCP_SYN_SENT] = TCP_CLOSE,
2004 [TCP_SYN_RECV] = TCP_FIN_WAIT1 | TCP_ACTION_FIN,
2005 [TCP_FIN_WAIT1] = TCP_FIN_WAIT1,
2006 [TCP_FIN_WAIT2] = TCP_FIN_WAIT2,
2007 [TCP_TIME_WAIT] = TCP_CLOSE,
2008 [TCP_CLOSE] = TCP_CLOSE,
2009 [TCP_CLOSE_WAIT] = TCP_LAST_ACK | TCP_ACTION_FIN,
2010 [TCP_LAST_ACK] = TCP_LAST_ACK,
2011 [TCP_LISTEN] = TCP_CLOSE,
2012 [TCP_CLOSING] = TCP_CLOSING,
2013 [TCP_NEW_SYN_RECV] = TCP_CLOSE, /* should not happen ! */
2014};
2015
2016static int tcp_close_state(struct sock *sk)
2017{
2018 int next = (int)new_state[sk->sk_state];
2019 int ns = next & TCP_STATE_MASK;
2020
2021 tcp_set_state(sk, ns);
2022
2023 return next & TCP_ACTION_FIN;
2024}
2025
2026/*
2027 * Shutdown the sending side of a connection. Much like close except
2028 * that we don't receive shut down or sock_set_flag(sk, SOCK_DEAD).
2029 */
2030
2031void tcp_shutdown(struct sock *sk, int how)
2032{
2033 /* We need to grab some memory, and put together a FIN,
2034 * and then put it into the queue to be sent.
2035 * Tim MacKenzie(tym@dibbler.cs.monash.edu.au) 4 Dec '92.
2036 */
2037 if (!(how & SEND_SHUTDOWN))
2038 return;
2039
2040 /* If we've already sent a FIN, or it's a closed state, skip this. */
2041 if ((1 << sk->sk_state) &
2042 (TCPF_ESTABLISHED | TCPF_SYN_SENT |
2043 TCPF_SYN_RECV | TCPF_CLOSE_WAIT)) {
2044 /* Clear out any half completed packets. FIN if needed. */
2045 if (tcp_close_state(sk))
2046 tcp_send_fin(sk);
2047 }
2048}
2049EXPORT_SYMBOL(tcp_shutdown);
2050
2051bool tcp_check_oom(struct sock *sk, int shift)
2052{
2053 bool too_many_orphans, out_of_socket_memory;
2054
2055 too_many_orphans = tcp_too_many_orphans(sk, shift);
2056 out_of_socket_memory = tcp_out_of_memory(sk);
2057
2058 if (too_many_orphans)
2059 net_info_ratelimited("too many orphaned sockets\n");
2060 if (out_of_socket_memory)
2061 net_info_ratelimited("out of memory -- consider tuning tcp_mem\n");
2062 return too_many_orphans || out_of_socket_memory;
2063}
2064
2065void tcp_close(struct sock *sk, long timeout)
2066{
2067 struct sk_buff *skb;
2068 int data_was_unread = 0;
2069 int state;
2070
2071 lock_sock(sk);
2072 sk->sk_shutdown = SHUTDOWN_MASK;
2073
2074 if (sk->sk_state == TCP_LISTEN) {
2075 tcp_set_state(sk, TCP_CLOSE);
2076
2077 /* Special case. */
2078 inet_csk_listen_stop(sk);
2079
2080 goto adjudge_to_death;
2081 }
2082
2083 /* We need to flush the recv. buffs. We do this only on the
2084 * descriptor close, not protocol-sourced closes, because the
2085 * reader process may not have drained the data yet!
2086 */
2087 while ((skb = __skb_dequeue(&sk->sk_receive_queue)) != NULL) {
2088 u32 len = TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq;
2089
2090 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
2091 len--;
2092 data_was_unread += len;
2093 __kfree_skb(skb);
2094 }
2095
2096 sk_mem_reclaim(sk);
2097
2098 /* If socket has been already reset (e.g. in tcp_reset()) - kill it. */
2099 if (sk->sk_state == TCP_CLOSE)
2100 goto adjudge_to_death;
2101
2102 /* As outlined in RFC 2525, section 2.17, we send a RST here because
2103 * data was lost. To witness the awful effects of the old behavior of
2104 * always doing a FIN, run an older 2.1.x kernel or 2.0.x, start a bulk
2105 * GET in an FTP client, suspend the process, wait for the client to
2106 * advertise a zero window, then kill -9 the FTP client, wheee...
2107 * Note: timeout is always zero in such a case.
2108 */
2109 if (unlikely(tcp_sk(sk)->repair)) {
2110 sk->sk_prot->disconnect(sk, 0);
2111 } else if (data_was_unread) {
2112 /* Unread data was tossed, zap the connection. */
2113 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONCLOSE);
2114 tcp_set_state(sk, TCP_CLOSE);
2115 tcp_send_active_reset(sk, sk->sk_allocation);
2116 } else if (sock_flag(sk, SOCK_LINGER) && !sk->sk_lingertime) {
2117 /* Check zero linger _after_ checking for unread data. */
2118 sk->sk_prot->disconnect(sk, 0);
2119 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONDATA);
2120 } else if (tcp_close_state(sk)) {
2121 /* We FIN if the application ate all the data before
2122 * zapping the connection.
2123 */
2124
2125 /* RED-PEN. Formally speaking, we have broken TCP state
2126 * machine. State transitions:
2127 *
2128 * TCP_ESTABLISHED -> TCP_FIN_WAIT1
2129 * TCP_SYN_RECV -> TCP_FIN_WAIT1 (forget it, it's impossible)
2130 * TCP_CLOSE_WAIT -> TCP_LAST_ACK
2131 *
2132 * are legal only when FIN has been sent (i.e. in window),
2133 * rather than queued out of window. Purists blame.
2134 *
2135 * F.e. "RFC state" is ESTABLISHED,
2136 * if Linux state is FIN-WAIT-1, but FIN is still not sent.
2137 *
2138 * The visible declinations are that sometimes
2139 * we enter time-wait state, when it is not required really
2140 * (harmless), do not send active resets, when they are
2141 * required by specs (TCP_ESTABLISHED, TCP_CLOSE_WAIT, when
2142 * they look as CLOSING or LAST_ACK for Linux)
2143 * Probably, I missed some more holelets.
2144 * --ANK
2145 * XXX (TFO) - To start off we don't support SYN+ACK+FIN
2146 * in a single packet! (May consider it later but will
2147 * probably need API support or TCP_CORK SYN-ACK until
2148 * data is written and socket is closed.)
2149 */
2150 tcp_send_fin(sk);
2151 }
2152
2153 sk_stream_wait_close(sk, timeout);
2154
2155adjudge_to_death:
2156 state = sk->sk_state;
2157 sock_hold(sk);
2158 sock_orphan(sk);
2159
2160 /* It is the last release_sock in its life. It will remove backlog. */
2161 release_sock(sk);
2162
2163
2164 /* Now socket is owned by kernel and we acquire BH lock
2165 to finish close. No need to check for user refs.
2166 */
2167 local_bh_disable();
2168 bh_lock_sock(sk);
2169 WARN_ON(sock_owned_by_user(sk));
2170
2171 percpu_counter_inc(sk->sk_prot->orphan_count);
2172
2173 /* Have we already been destroyed by a softirq or backlog? */
2174 if (state != TCP_CLOSE && sk->sk_state == TCP_CLOSE)
2175 goto out;
2176
2177 /* This is a (useful) BSD violating of the RFC. There is a
2178 * problem with TCP as specified in that the other end could
2179 * keep a socket open forever with no application left this end.
2180 * We use a 1 minute timeout (about the same as BSD) then kill
2181 * our end. If they send after that then tough - BUT: long enough
2182 * that we won't make the old 4*rto = almost no time - whoops
2183 * reset mistake.
2184 *
2185 * Nope, it was not mistake. It is really desired behaviour
2186 * f.e. on http servers, when such sockets are useless, but
2187 * consume significant resources. Let's do it with special
2188 * linger2 option. --ANK
2189 */
2190
2191 if (sk->sk_state == TCP_FIN_WAIT2) {
2192 struct tcp_sock *tp = tcp_sk(sk);
2193 if (tp->linger2 < 0) {
2194 tcp_set_state(sk, TCP_CLOSE);
2195 tcp_send_active_reset(sk, GFP_ATOMIC);
2196 __NET_INC_STATS(sock_net(sk),
2197 LINUX_MIB_TCPABORTONLINGER);
2198 } else {
2199 const int tmo = tcp_fin_time(sk);
2200
2201 if (tmo > TCP_TIMEWAIT_LEN) {
2202 inet_csk_reset_keepalive_timer(sk,
2203 tmo - TCP_TIMEWAIT_LEN);
2204 } else {
2205 tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
2206 goto out;
2207 }
2208 }
2209 }
2210 if (sk->sk_state != TCP_CLOSE) {
2211 sk_mem_reclaim(sk);
2212 if (tcp_check_oom(sk, 0)) {
2213 tcp_set_state(sk, TCP_CLOSE);
2214 tcp_send_active_reset(sk, GFP_ATOMIC);
2215 __NET_INC_STATS(sock_net(sk),
2216 LINUX_MIB_TCPABORTONMEMORY);
2217 }
2218 }
2219
2220 if (sk->sk_state == TCP_CLOSE) {
2221 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
2222 /* We could get here with a non-NULL req if the socket is
2223 * aborted (e.g., closed with unread data) before 3WHS
2224 * finishes.
2225 */
2226 if (req)
2227 reqsk_fastopen_remove(sk, req, false);
2228 inet_csk_destroy_sock(sk);
2229 }
2230 /* Otherwise, socket is reprieved until protocol close. */
2231
2232out:
2233 bh_unlock_sock(sk);
2234 local_bh_enable();
2235 sock_put(sk);
2236}
2237EXPORT_SYMBOL(tcp_close);
2238
2239/* These states need RST on ABORT according to RFC793 */
2240
2241static inline bool tcp_need_reset(int state)
2242{
2243 return (1 << state) &
2244 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT | TCPF_FIN_WAIT1 |
2245 TCPF_FIN_WAIT2 | TCPF_SYN_RECV);
2246}
2247
2248int tcp_disconnect(struct sock *sk, int flags)
2249{
2250 struct inet_sock *inet = inet_sk(sk);
2251 struct inet_connection_sock *icsk = inet_csk(sk);
2252 struct tcp_sock *tp = tcp_sk(sk);
2253 int err = 0;
2254 int old_state = sk->sk_state;
2255
2256 if (old_state != TCP_CLOSE)
2257 tcp_set_state(sk, TCP_CLOSE);
2258
2259 /* ABORT function of RFC793 */
2260 if (old_state == TCP_LISTEN) {
2261 inet_csk_listen_stop(sk);
2262 } else if (unlikely(tp->repair)) {
2263 sk->sk_err = ECONNABORTED;
2264 } else if (tcp_need_reset(old_state) ||
2265 (tp->snd_nxt != tp->write_seq &&
2266 (1 << old_state) & (TCPF_CLOSING | TCPF_LAST_ACK))) {
2267 /* The last check adjusts for discrepancy of Linux wrt. RFC
2268 * states
2269 */
2270 tcp_send_active_reset(sk, gfp_any());
2271 sk->sk_err = ECONNRESET;
2272 } else if (old_state == TCP_SYN_SENT)
2273 sk->sk_err = ECONNRESET;
2274
2275 tcp_clear_xmit_timers(sk);
2276 __skb_queue_purge(&sk->sk_receive_queue);
2277 tcp_write_queue_purge(sk);
2278 skb_rbtree_purge(&tp->out_of_order_queue);
2279
2280 inet->inet_dport = 0;
2281
2282 if (!(sk->sk_userlocks & SOCK_BINDADDR_LOCK))
2283 inet_reset_saddr(sk);
2284
2285 sk->sk_shutdown = 0;
2286 sock_reset_flag(sk, SOCK_DONE);
2287 tp->srtt_us = 0;
2288 tp->write_seq += tp->max_window + 2;
2289 if (tp->write_seq == 0)
2290 tp->write_seq = 1;
2291 icsk->icsk_backoff = 0;
2292 tp->snd_cwnd = 2;
2293 icsk->icsk_probes_out = 0;
2294 tp->packets_out = 0;
2295 tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
2296 tp->snd_cwnd_cnt = 0;
2297 tp->window_clamp = 0;
2298 tcp_set_ca_state(sk, TCP_CA_Open);
2299 tcp_clear_retrans(tp);
2300 inet_csk_delack_init(sk);
2301 tcp_init_send_head(sk);
2302 memset(&tp->rx_opt, 0, sizeof(tp->rx_opt));
2303 __sk_dst_reset(sk);
2304
2305 WARN_ON(inet->inet_num && !icsk->icsk_bind_hash);
2306
2307 sk->sk_error_report(sk);
2308 return err;
2309}
2310EXPORT_SYMBOL(tcp_disconnect);
2311
2312static inline bool tcp_can_repair_sock(const struct sock *sk)
2313{
2314 return ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN) &&
2315 (sk->sk_state != TCP_LISTEN);
2316}
2317
2318static int tcp_repair_set_window(struct tcp_sock *tp, char __user *optbuf, int len)
2319{
2320 struct tcp_repair_window opt;
2321
2322 if (!tp->repair)
2323 return -EPERM;
2324
2325 if (len != sizeof(opt))
2326 return -EINVAL;
2327
2328 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2329 return -EFAULT;
2330
2331 if (opt.max_window < opt.snd_wnd)
2332 return -EINVAL;
2333
2334 if (after(opt.snd_wl1, tp->rcv_nxt + opt.rcv_wnd))
2335 return -EINVAL;
2336
2337 if (after(opt.rcv_wup, tp->rcv_nxt))
2338 return -EINVAL;
2339
2340 tp->snd_wl1 = opt.snd_wl1;
2341 tp->snd_wnd = opt.snd_wnd;
2342 tp->max_window = opt.max_window;
2343
2344 tp->rcv_wnd = opt.rcv_wnd;
2345 tp->rcv_wup = opt.rcv_wup;
2346
2347 return 0;
2348}
2349
2350static int tcp_repair_options_est(struct tcp_sock *tp,
2351 struct tcp_repair_opt __user *optbuf, unsigned int len)
2352{
2353 struct tcp_repair_opt opt;
2354
2355 while (len >= sizeof(opt)) {
2356 if (copy_from_user(&opt, optbuf, sizeof(opt)))
2357 return -EFAULT;
2358
2359 optbuf++;
2360 len -= sizeof(opt);
2361
2362 switch (opt.opt_code) {
2363 case TCPOPT_MSS:
2364 tp->rx_opt.mss_clamp = opt.opt_val;
2365 break;
2366 case TCPOPT_WINDOW:
2367 {
2368 u16 snd_wscale = opt.opt_val & 0xFFFF;
2369 u16 rcv_wscale = opt.opt_val >> 16;
2370
2371 if (snd_wscale > 14 || rcv_wscale > 14)
2372 return -EFBIG;
2373
2374 tp->rx_opt.snd_wscale = snd_wscale;
2375 tp->rx_opt.rcv_wscale = rcv_wscale;
2376 tp->rx_opt.wscale_ok = 1;
2377 }
2378 break;
2379 case TCPOPT_SACK_PERM:
2380 if (opt.opt_val != 0)
2381 return -EINVAL;
2382
2383 tp->rx_opt.sack_ok |= TCP_SACK_SEEN;
2384 if (sysctl_tcp_fack)
2385 tcp_enable_fack(tp);
2386 break;
2387 case TCPOPT_TIMESTAMP:
2388 if (opt.opt_val != 0)
2389 return -EINVAL;
2390
2391 tp->rx_opt.tstamp_ok = 1;
2392 break;
2393 }
2394 }
2395
2396 return 0;
2397}
2398
2399/*
2400 * Socket option code for TCP.
2401 */
2402static int do_tcp_setsockopt(struct sock *sk, int level,
2403 int optname, char __user *optval, unsigned int optlen)
2404{
2405 struct tcp_sock *tp = tcp_sk(sk);
2406 struct inet_connection_sock *icsk = inet_csk(sk);
2407 struct net *net = sock_net(sk);
2408 int val;
2409 int err = 0;
2410
2411 /* These are data/string values, all the others are ints */
2412 switch (optname) {
2413 case TCP_CONGESTION: {
2414 char name[TCP_CA_NAME_MAX];
2415
2416 if (optlen < 1)
2417 return -EINVAL;
2418
2419 val = strncpy_from_user(name, optval,
2420 min_t(long, TCP_CA_NAME_MAX-1, optlen));
2421 if (val < 0)
2422 return -EFAULT;
2423 name[val] = 0;
2424
2425 lock_sock(sk);
2426 err = tcp_set_congestion_control(sk, name);
2427 release_sock(sk);
2428 return err;
2429 }
2430 default:
2431 /* fallthru */
2432 break;
2433 }
2434
2435 if (optlen < sizeof(int))
2436 return -EINVAL;
2437
2438 if (get_user(val, (int __user *)optval))
2439 return -EFAULT;
2440
2441 lock_sock(sk);
2442
2443 switch (optname) {
2444 case TCP_MAXSEG:
2445 /* Values greater than interface MTU won't take effect. However
2446 * at the point when this call is done we typically don't yet
2447 * know which interface is going to be used */
2448 if (val < TCP_MIN_MSS || val > MAX_TCP_WINDOW) {
2449 err = -EINVAL;
2450 break;
2451 }
2452 tp->rx_opt.user_mss = val;
2453 break;
2454
2455 case TCP_NODELAY:
2456 if (val) {
2457 /* TCP_NODELAY is weaker than TCP_CORK, so that
2458 * this option on corked socket is remembered, but
2459 * it is not activated until cork is cleared.
2460 *
2461 * However, when TCP_NODELAY is set we make
2462 * an explicit push, which overrides even TCP_CORK
2463 * for currently queued segments.
2464 */
2465 tp->nonagle |= TCP_NAGLE_OFF|TCP_NAGLE_PUSH;
2466 tcp_push_pending_frames(sk);
2467 } else {
2468 tp->nonagle &= ~TCP_NAGLE_OFF;
2469 }
2470 break;
2471
2472 case TCP_THIN_LINEAR_TIMEOUTS:
2473 if (val < 0 || val > 1)
2474 err = -EINVAL;
2475 else
2476 tp->thin_lto = val;
2477 break;
2478
2479 case TCP_THIN_DUPACK:
2480 if (val < 0 || val > 1)
2481 err = -EINVAL;
2482 else {
2483 tp->thin_dupack = val;
2484 if (tp->thin_dupack)
2485 tcp_disable_early_retrans(tp);
2486 }
2487 break;
2488
2489 case TCP_REPAIR:
2490 if (!tcp_can_repair_sock(sk))
2491 err = -EPERM;
2492 else if (val == 1) {
2493 tp->repair = 1;
2494 sk->sk_reuse = SK_FORCE_REUSE;
2495 tp->repair_queue = TCP_NO_QUEUE;
2496 } else if (val == 0) {
2497 tp->repair = 0;
2498 sk->sk_reuse = SK_NO_REUSE;
2499 tcp_send_window_probe(sk);
2500 } else
2501 err = -EINVAL;
2502
2503 break;
2504
2505 case TCP_REPAIR_QUEUE:
2506 if (!tp->repair)
2507 err = -EPERM;
2508 else if (val < TCP_QUEUES_NR)
2509 tp->repair_queue = val;
2510 else
2511 err = -EINVAL;
2512 break;
2513
2514 case TCP_QUEUE_SEQ:
2515 if (sk->sk_state != TCP_CLOSE)
2516 err = -EPERM;
2517 else if (tp->repair_queue == TCP_SEND_QUEUE)
2518 tp->write_seq = val;
2519 else if (tp->repair_queue == TCP_RECV_QUEUE)
2520 tp->rcv_nxt = val;
2521 else
2522 err = -EINVAL;
2523 break;
2524
2525 case TCP_REPAIR_OPTIONS:
2526 if (!tp->repair)
2527 err = -EINVAL;
2528 else if (sk->sk_state == TCP_ESTABLISHED)
2529 err = tcp_repair_options_est(tp,
2530 (struct tcp_repair_opt __user *)optval,
2531 optlen);
2532 else
2533 err = -EPERM;
2534 break;
2535
2536 case TCP_CORK:
2537 /* When set indicates to always queue non-full frames.
2538 * Later the user clears this option and we transmit
2539 * any pending partial frames in the queue. This is
2540 * meant to be used alongside sendfile() to get properly
2541 * filled frames when the user (for example) must write
2542 * out headers with a write() call first and then use
2543 * sendfile to send out the data parts.
2544 *
2545 * TCP_CORK can be set together with TCP_NODELAY and it is
2546 * stronger than TCP_NODELAY.
2547 */
2548 if (val) {
2549 tp->nonagle |= TCP_NAGLE_CORK;
2550 } else {
2551 tp->nonagle &= ~TCP_NAGLE_CORK;
2552 if (tp->nonagle&TCP_NAGLE_OFF)
2553 tp->nonagle |= TCP_NAGLE_PUSH;
2554 tcp_push_pending_frames(sk);
2555 }
2556 break;
2557
2558 case TCP_KEEPIDLE:
2559 if (val < 1 || val > MAX_TCP_KEEPIDLE)
2560 err = -EINVAL;
2561 else {
2562 tp->keepalive_time = val * HZ;
2563 if (sock_flag(sk, SOCK_KEEPOPEN) &&
2564 !((1 << sk->sk_state) &
2565 (TCPF_CLOSE | TCPF_LISTEN))) {
2566 u32 elapsed = keepalive_time_elapsed(tp);
2567 if (tp->keepalive_time > elapsed)
2568 elapsed = tp->keepalive_time - elapsed;
2569 else
2570 elapsed = 0;
2571 inet_csk_reset_keepalive_timer(sk, elapsed);
2572 }
2573 }
2574 break;
2575 case TCP_KEEPINTVL:
2576 if (val < 1 || val > MAX_TCP_KEEPINTVL)
2577 err = -EINVAL;
2578 else
2579 tp->keepalive_intvl = val * HZ;
2580 break;
2581 case TCP_KEEPCNT:
2582 if (val < 1 || val > MAX_TCP_KEEPCNT)
2583 err = -EINVAL;
2584 else
2585 tp->keepalive_probes = val;
2586 break;
2587 case TCP_SYNCNT:
2588 if (val < 1 || val > MAX_TCP_SYNCNT)
2589 err = -EINVAL;
2590 else
2591 icsk->icsk_syn_retries = val;
2592 break;
2593
2594 case TCP_SAVE_SYN:
2595 if (val < 0 || val > 1)
2596 err = -EINVAL;
2597 else
2598 tp->save_syn = val;
2599 break;
2600
2601 case TCP_LINGER2:
2602 if (val < 0)
2603 tp->linger2 = -1;
2604 else if (val > net->ipv4.sysctl_tcp_fin_timeout / HZ)
2605 tp->linger2 = 0;
2606 else
2607 tp->linger2 = val * HZ;
2608 break;
2609
2610 case TCP_DEFER_ACCEPT:
2611 /* Translate value in seconds to number of retransmits */
2612 icsk->icsk_accept_queue.rskq_defer_accept =
2613 secs_to_retrans(val, TCP_TIMEOUT_INIT / HZ,
2614 TCP_RTO_MAX / HZ);
2615 break;
2616
2617 case TCP_WINDOW_CLAMP:
2618 if (!val) {
2619 if (sk->sk_state != TCP_CLOSE) {
2620 err = -EINVAL;
2621 break;
2622 }
2623 tp->window_clamp = 0;
2624 } else
2625 tp->window_clamp = val < SOCK_MIN_RCVBUF / 2 ?
2626 SOCK_MIN_RCVBUF / 2 : val;
2627 break;
2628
2629 case TCP_QUICKACK:
2630 if (!val) {
2631 icsk->icsk_ack.pingpong = 1;
2632 } else {
2633 icsk->icsk_ack.pingpong = 0;
2634 if ((1 << sk->sk_state) &
2635 (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT) &&
2636 inet_csk_ack_scheduled(sk)) {
2637 icsk->icsk_ack.pending |= ICSK_ACK_PUSHED;
2638 tcp_cleanup_rbuf(sk, 1);
2639 if (!(val & 1))
2640 icsk->icsk_ack.pingpong = 1;
2641 }
2642 }
2643 break;
2644
2645#ifdef CONFIG_TCP_MD5SIG
2646 case TCP_MD5SIG:
2647 /* Read the IP->Key mappings from userspace */
2648 err = tp->af_specific->md5_parse(sk, optval, optlen);
2649 break;
2650#endif
2651 case TCP_USER_TIMEOUT:
2652 /* Cap the max time in ms TCP will retry or probe the window
2653 * before giving up and aborting (ETIMEDOUT) a connection.
2654 */
2655 if (val < 0)
2656 err = -EINVAL;
2657 else
2658 icsk->icsk_user_timeout = msecs_to_jiffies(val);
2659 break;
2660
2661 case TCP_FASTOPEN:
2662 if (val >= 0 && ((1 << sk->sk_state) & (TCPF_CLOSE |
2663 TCPF_LISTEN))) {
2664 tcp_fastopen_init_key_once(true);
2665
2666 fastopen_queue_tune(sk, val);
2667 } else {
2668 err = -EINVAL;
2669 }
2670 break;
2671 case TCP_TIMESTAMP:
2672 if (!tp->repair)
2673 err = -EPERM;
2674 else
2675 tp->tsoffset = val - tcp_time_stamp;
2676 break;
2677 case TCP_REPAIR_WINDOW:
2678 err = tcp_repair_set_window(tp, optval, optlen);
2679 break;
2680 case TCP_NOTSENT_LOWAT:
2681 tp->notsent_lowat = val;
2682 sk->sk_write_space(sk);
2683 break;
2684 default:
2685 err = -ENOPROTOOPT;
2686 break;
2687 }
2688
2689 release_sock(sk);
2690 return err;
2691}
2692
2693int tcp_setsockopt(struct sock *sk, int level, int optname, char __user *optval,
2694 unsigned int optlen)
2695{
2696 const struct inet_connection_sock *icsk = inet_csk(sk);
2697
2698 if (level != SOL_TCP)
2699 return icsk->icsk_af_ops->setsockopt(sk, level, optname,
2700 optval, optlen);
2701 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2702}
2703EXPORT_SYMBOL(tcp_setsockopt);
2704
2705#ifdef CONFIG_COMPAT
2706int compat_tcp_setsockopt(struct sock *sk, int level, int optname,
2707 char __user *optval, unsigned int optlen)
2708{
2709 if (level != SOL_TCP)
2710 return inet_csk_compat_setsockopt(sk, level, optname,
2711 optval, optlen);
2712 return do_tcp_setsockopt(sk, level, optname, optval, optlen);
2713}
2714EXPORT_SYMBOL(compat_tcp_setsockopt);
2715#endif
2716
2717static void tcp_get_info_chrono_stats(const struct tcp_sock *tp,
2718 struct tcp_info *info)
2719{
2720 u64 stats[__TCP_CHRONO_MAX], total = 0;
2721 enum tcp_chrono i;
2722
2723 for (i = TCP_CHRONO_BUSY; i < __TCP_CHRONO_MAX; ++i) {
2724 stats[i] = tp->chrono_stat[i - 1];
2725 if (i == tp->chrono_type)
2726 stats[i] += tcp_time_stamp - tp->chrono_start;
2727 stats[i] *= USEC_PER_SEC / HZ;
2728 total += stats[i];
2729 }
2730
2731 info->tcpi_busy_time = total;
2732 info->tcpi_rwnd_limited = stats[TCP_CHRONO_RWND_LIMITED];
2733 info->tcpi_sndbuf_limited = stats[TCP_CHRONO_SNDBUF_LIMITED];
2734}
2735
2736/* Return information about state of tcp endpoint in API format. */
2737void tcp_get_info(struct sock *sk, struct tcp_info *info)
2738{
2739 const struct tcp_sock *tp = tcp_sk(sk); /* iff sk_type == SOCK_STREAM */
2740 const struct inet_connection_sock *icsk = inet_csk(sk);
2741 u32 now = tcp_time_stamp, intv;
2742 u64 rate64;
2743 bool slow;
2744 u32 rate;
2745
2746 memset(info, 0, sizeof(*info));
2747 if (sk->sk_type != SOCK_STREAM)
2748 return;
2749
2750 info->tcpi_state = sk_state_load(sk);
2751
2752 /* Report meaningful fields for all TCP states, including listeners */
2753 rate = READ_ONCE(sk->sk_pacing_rate);
2754 rate64 = rate != ~0U ? rate : ~0ULL;
2755 info->tcpi_pacing_rate = rate64;
2756
2757 rate = READ_ONCE(sk->sk_max_pacing_rate);
2758 rate64 = rate != ~0U ? rate : ~0ULL;
2759 info->tcpi_max_pacing_rate = rate64;
2760
2761 info->tcpi_reordering = tp->reordering;
2762 info->tcpi_snd_cwnd = tp->snd_cwnd;
2763
2764 if (info->tcpi_state == TCP_LISTEN) {
2765 /* listeners aliased fields :
2766 * tcpi_unacked -> Number of children ready for accept()
2767 * tcpi_sacked -> max backlog
2768 */
2769 info->tcpi_unacked = sk->sk_ack_backlog;
2770 info->tcpi_sacked = sk->sk_max_ack_backlog;
2771 return;
2772 }
2773 info->tcpi_ca_state = icsk->icsk_ca_state;
2774 info->tcpi_retransmits = icsk->icsk_retransmits;
2775 info->tcpi_probes = icsk->icsk_probes_out;
2776 info->tcpi_backoff = icsk->icsk_backoff;
2777
2778 if (tp->rx_opt.tstamp_ok)
2779 info->tcpi_options |= TCPI_OPT_TIMESTAMPS;
2780 if (tcp_is_sack(tp))
2781 info->tcpi_options |= TCPI_OPT_SACK;
2782 if (tp->rx_opt.wscale_ok) {
2783 info->tcpi_options |= TCPI_OPT_WSCALE;
2784 info->tcpi_snd_wscale = tp->rx_opt.snd_wscale;
2785 info->tcpi_rcv_wscale = tp->rx_opt.rcv_wscale;
2786 }
2787
2788 if (tp->ecn_flags & TCP_ECN_OK)
2789 info->tcpi_options |= TCPI_OPT_ECN;
2790 if (tp->ecn_flags & TCP_ECN_SEEN)
2791 info->tcpi_options |= TCPI_OPT_ECN_SEEN;
2792 if (tp->syn_data_acked)
2793 info->tcpi_options |= TCPI_OPT_SYN_DATA;
2794
2795 info->tcpi_rto = jiffies_to_usecs(icsk->icsk_rto);
2796 info->tcpi_ato = jiffies_to_usecs(icsk->icsk_ack.ato);
2797 info->tcpi_snd_mss = tp->mss_cache;
2798 info->tcpi_rcv_mss = icsk->icsk_ack.rcv_mss;
2799
2800 info->tcpi_unacked = tp->packets_out;
2801 info->tcpi_sacked = tp->sacked_out;
2802
2803 info->tcpi_lost = tp->lost_out;
2804 info->tcpi_retrans = tp->retrans_out;
2805 info->tcpi_fackets = tp->fackets_out;
2806
2807 info->tcpi_last_data_sent = jiffies_to_msecs(now - tp->lsndtime);
2808 info->tcpi_last_data_recv = jiffies_to_msecs(now - icsk->icsk_ack.lrcvtime);
2809 info->tcpi_last_ack_recv = jiffies_to_msecs(now - tp->rcv_tstamp);
2810
2811 info->tcpi_pmtu = icsk->icsk_pmtu_cookie;
2812 info->tcpi_rcv_ssthresh = tp->rcv_ssthresh;
2813 info->tcpi_rtt = tp->srtt_us >> 3;
2814 info->tcpi_rttvar = tp->mdev_us >> 2;
2815 info->tcpi_snd_ssthresh = tp->snd_ssthresh;
2816 info->tcpi_advmss = tp->advmss;
2817
2818 info->tcpi_rcv_rtt = jiffies_to_usecs(tp->rcv_rtt_est.rtt)>>3;
2819 info->tcpi_rcv_space = tp->rcvq_space.space;
2820
2821 info->tcpi_total_retrans = tp->total_retrans;
2822
2823 slow = lock_sock_fast(sk);
2824
2825 info->tcpi_bytes_acked = tp->bytes_acked;
2826 info->tcpi_bytes_received = tp->bytes_received;
2827 info->tcpi_notsent_bytes = max_t(int, 0, tp->write_seq - tp->snd_nxt);
2828 tcp_get_info_chrono_stats(tp, info);
2829
2830 unlock_sock_fast(sk, slow);
2831
2832 info->tcpi_segs_out = tp->segs_out;
2833 info->tcpi_segs_in = tp->segs_in;
2834
2835 info->tcpi_min_rtt = tcp_min_rtt(tp);
2836 info->tcpi_data_segs_in = tp->data_segs_in;
2837 info->tcpi_data_segs_out = tp->data_segs_out;
2838
2839 info->tcpi_delivery_rate_app_limited = tp->rate_app_limited ? 1 : 0;
2840 rate = READ_ONCE(tp->rate_delivered);
2841 intv = READ_ONCE(tp->rate_interval_us);
2842 if (rate && intv) {
2843 rate64 = (u64)rate * tp->mss_cache * USEC_PER_SEC;
2844 do_div(rate64, intv);
2845 info->tcpi_delivery_rate = rate64;
2846 }
2847}
2848EXPORT_SYMBOL_GPL(tcp_get_info);
2849
2850struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk)
2851{
2852 const struct tcp_sock *tp = tcp_sk(sk);
2853 struct sk_buff *stats;
2854 struct tcp_info info;
2855
2856 stats = alloc_skb(3 * nla_total_size_64bit(sizeof(u64)), GFP_ATOMIC);
2857 if (!stats)
2858 return NULL;
2859
2860 tcp_get_info_chrono_stats(tp, &info);
2861 nla_put_u64_64bit(stats, TCP_NLA_BUSY,
2862 info.tcpi_busy_time, TCP_NLA_PAD);
2863 nla_put_u64_64bit(stats, TCP_NLA_RWND_LIMITED,
2864 info.tcpi_rwnd_limited, TCP_NLA_PAD);
2865 nla_put_u64_64bit(stats, TCP_NLA_SNDBUF_LIMITED,
2866 info.tcpi_sndbuf_limited, TCP_NLA_PAD);
2867 return stats;
2868}
2869
2870static int do_tcp_getsockopt(struct sock *sk, int level,
2871 int optname, char __user *optval, int __user *optlen)
2872{
2873 struct inet_connection_sock *icsk = inet_csk(sk);
2874 struct tcp_sock *tp = tcp_sk(sk);
2875 struct net *net = sock_net(sk);
2876 int val, len;
2877
2878 if (get_user(len, optlen))
2879 return -EFAULT;
2880
2881 len = min_t(unsigned int, len, sizeof(int));
2882
2883 if (len < 0)
2884 return -EINVAL;
2885
2886 switch (optname) {
2887 case TCP_MAXSEG:
2888 val = tp->mss_cache;
2889 if (!val && ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
2890 val = tp->rx_opt.user_mss;
2891 if (tp->repair)
2892 val = tp->rx_opt.mss_clamp;
2893 break;
2894 case TCP_NODELAY:
2895 val = !!(tp->nonagle&TCP_NAGLE_OFF);
2896 break;
2897 case TCP_CORK:
2898 val = !!(tp->nonagle&TCP_NAGLE_CORK);
2899 break;
2900 case TCP_KEEPIDLE:
2901 val = keepalive_time_when(tp) / HZ;
2902 break;
2903 case TCP_KEEPINTVL:
2904 val = keepalive_intvl_when(tp) / HZ;
2905 break;
2906 case TCP_KEEPCNT:
2907 val = keepalive_probes(tp);
2908 break;
2909 case TCP_SYNCNT:
2910 val = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
2911 break;
2912 case TCP_LINGER2:
2913 val = tp->linger2;
2914 if (val >= 0)
2915 val = (val ? : net->ipv4.sysctl_tcp_fin_timeout) / HZ;
2916 break;
2917 case TCP_DEFER_ACCEPT:
2918 val = retrans_to_secs(icsk->icsk_accept_queue.rskq_defer_accept,
2919 TCP_TIMEOUT_INIT / HZ, TCP_RTO_MAX / HZ);
2920 break;
2921 case TCP_WINDOW_CLAMP:
2922 val = tp->window_clamp;
2923 break;
2924 case TCP_INFO: {
2925 struct tcp_info info;
2926
2927 if (get_user(len, optlen))
2928 return -EFAULT;
2929
2930 tcp_get_info(sk, &info);
2931
2932 len = min_t(unsigned int, len, sizeof(info));
2933 if (put_user(len, optlen))
2934 return -EFAULT;
2935 if (copy_to_user(optval, &info, len))
2936 return -EFAULT;
2937 return 0;
2938 }
2939 case TCP_CC_INFO: {
2940 const struct tcp_congestion_ops *ca_ops;
2941 union tcp_cc_info info;
2942 size_t sz = 0;
2943 int attr;
2944
2945 if (get_user(len, optlen))
2946 return -EFAULT;
2947
2948 ca_ops = icsk->icsk_ca_ops;
2949 if (ca_ops && ca_ops->get_info)
2950 sz = ca_ops->get_info(sk, ~0U, &attr, &info);
2951
2952 len = min_t(unsigned int, len, sz);
2953 if (put_user(len, optlen))
2954 return -EFAULT;
2955 if (copy_to_user(optval, &info, len))
2956 return -EFAULT;
2957 return 0;
2958 }
2959 case TCP_QUICKACK:
2960 val = !icsk->icsk_ack.pingpong;
2961 break;
2962
2963 case TCP_CONGESTION:
2964 if (get_user(len, optlen))
2965 return -EFAULT;
2966 len = min_t(unsigned int, len, TCP_CA_NAME_MAX);
2967 if (put_user(len, optlen))
2968 return -EFAULT;
2969 if (copy_to_user(optval, icsk->icsk_ca_ops->name, len))
2970 return -EFAULT;
2971 return 0;
2972
2973 case TCP_THIN_LINEAR_TIMEOUTS:
2974 val = tp->thin_lto;
2975 break;
2976 case TCP_THIN_DUPACK:
2977 val = tp->thin_dupack;
2978 break;
2979
2980 case TCP_REPAIR:
2981 val = tp->repair;
2982 break;
2983
2984 case TCP_REPAIR_QUEUE:
2985 if (tp->repair)
2986 val = tp->repair_queue;
2987 else
2988 return -EINVAL;
2989 break;
2990
2991 case TCP_REPAIR_WINDOW: {
2992 struct tcp_repair_window opt;
2993
2994 if (get_user(len, optlen))
2995 return -EFAULT;
2996
2997 if (len != sizeof(opt))
2998 return -EINVAL;
2999
3000 if (!tp->repair)
3001 return -EPERM;
3002
3003 opt.snd_wl1 = tp->snd_wl1;
3004 opt.snd_wnd = tp->snd_wnd;
3005 opt.max_window = tp->max_window;
3006 opt.rcv_wnd = tp->rcv_wnd;
3007 opt.rcv_wup = tp->rcv_wup;
3008
3009 if (copy_to_user(optval, &opt, len))
3010 return -EFAULT;
3011 return 0;
3012 }
3013 case TCP_QUEUE_SEQ:
3014 if (tp->repair_queue == TCP_SEND_QUEUE)
3015 val = tp->write_seq;
3016 else if (tp->repair_queue == TCP_RECV_QUEUE)
3017 val = tp->rcv_nxt;
3018 else
3019 return -EINVAL;
3020 break;
3021
3022 case TCP_USER_TIMEOUT:
3023 val = jiffies_to_msecs(icsk->icsk_user_timeout);
3024 break;
3025
3026 case TCP_FASTOPEN:
3027 val = icsk->icsk_accept_queue.fastopenq.max_qlen;
3028 break;
3029
3030 case TCP_TIMESTAMP:
3031 val = tcp_time_stamp + tp->tsoffset;
3032 break;
3033 case TCP_NOTSENT_LOWAT:
3034 val = tp->notsent_lowat;
3035 break;
3036 case TCP_SAVE_SYN:
3037 val = tp->save_syn;
3038 break;
3039 case TCP_SAVED_SYN: {
3040 if (get_user(len, optlen))
3041 return -EFAULT;
3042
3043 lock_sock(sk);
3044 if (tp->saved_syn) {
3045 if (len < tp->saved_syn[0]) {
3046 if (put_user(tp->saved_syn[0], optlen)) {
3047 release_sock(sk);
3048 return -EFAULT;
3049 }
3050 release_sock(sk);
3051 return -EINVAL;
3052 }
3053 len = tp->saved_syn[0];
3054 if (put_user(len, optlen)) {
3055 release_sock(sk);
3056 return -EFAULT;
3057 }
3058 if (copy_to_user(optval, tp->saved_syn + 1, len)) {
3059 release_sock(sk);
3060 return -EFAULT;
3061 }
3062 tcp_saved_syn_free(tp);
3063 release_sock(sk);
3064 } else {
3065 release_sock(sk);
3066 len = 0;
3067 if (put_user(len, optlen))
3068 return -EFAULT;
3069 }
3070 return 0;
3071 }
3072 default:
3073 return -ENOPROTOOPT;
3074 }
3075
3076 if (put_user(len, optlen))
3077 return -EFAULT;
3078 if (copy_to_user(optval, &val, len))
3079 return -EFAULT;
3080 return 0;
3081}
3082
3083int tcp_getsockopt(struct sock *sk, int level, int optname, char __user *optval,
3084 int __user *optlen)
3085{
3086 struct inet_connection_sock *icsk = inet_csk(sk);
3087
3088 if (level != SOL_TCP)
3089 return icsk->icsk_af_ops->getsockopt(sk, level, optname,
3090 optval, optlen);
3091 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3092}
3093EXPORT_SYMBOL(tcp_getsockopt);
3094
3095#ifdef CONFIG_COMPAT
3096int compat_tcp_getsockopt(struct sock *sk, int level, int optname,
3097 char __user *optval, int __user *optlen)
3098{
3099 if (level != SOL_TCP)
3100 return inet_csk_compat_getsockopt(sk, level, optname,
3101 optval, optlen);
3102 return do_tcp_getsockopt(sk, level, optname, optval, optlen);
3103}
3104EXPORT_SYMBOL(compat_tcp_getsockopt);
3105#endif
3106
3107#ifdef CONFIG_TCP_MD5SIG
3108static DEFINE_PER_CPU(struct tcp_md5sig_pool, tcp_md5sig_pool);
3109static DEFINE_MUTEX(tcp_md5sig_mutex);
3110static bool tcp_md5sig_pool_populated = false;
3111
3112static void __tcp_alloc_md5sig_pool(void)
3113{
3114 struct crypto_ahash *hash;
3115 int cpu;
3116
3117 hash = crypto_alloc_ahash("md5", 0, CRYPTO_ALG_ASYNC);
3118 if (IS_ERR(hash))
3119 return;
3120
3121 for_each_possible_cpu(cpu) {
3122 void *scratch = per_cpu(tcp_md5sig_pool, cpu).scratch;
3123 struct ahash_request *req;
3124
3125 if (!scratch) {
3126 scratch = kmalloc_node(sizeof(union tcp_md5sum_block) +
3127 sizeof(struct tcphdr),
3128 GFP_KERNEL,
3129 cpu_to_node(cpu));
3130 if (!scratch)
3131 return;
3132 per_cpu(tcp_md5sig_pool, cpu).scratch = scratch;
3133 }
3134 if (per_cpu(tcp_md5sig_pool, cpu).md5_req)
3135 continue;
3136
3137 req = ahash_request_alloc(hash, GFP_KERNEL);
3138 if (!req)
3139 return;
3140
3141 ahash_request_set_callback(req, 0, NULL, NULL);
3142
3143 per_cpu(tcp_md5sig_pool, cpu).md5_req = req;
3144 }
3145 /* before setting tcp_md5sig_pool_populated, we must commit all writes
3146 * to memory. See smp_rmb() in tcp_get_md5sig_pool()
3147 */
3148 smp_wmb();
3149 tcp_md5sig_pool_populated = true;
3150}
3151
3152bool tcp_alloc_md5sig_pool(void)
3153{
3154 if (unlikely(!tcp_md5sig_pool_populated)) {
3155 mutex_lock(&tcp_md5sig_mutex);
3156
3157 if (!tcp_md5sig_pool_populated)
3158 __tcp_alloc_md5sig_pool();
3159
3160 mutex_unlock(&tcp_md5sig_mutex);
3161 }
3162 return tcp_md5sig_pool_populated;
3163}
3164EXPORT_SYMBOL(tcp_alloc_md5sig_pool);
3165
3166
3167/**
3168 * tcp_get_md5sig_pool - get md5sig_pool for this user
3169 *
3170 * We use percpu structure, so if we succeed, we exit with preemption
3171 * and BH disabled, to make sure another thread or softirq handling
3172 * wont try to get same context.
3173 */
3174struct tcp_md5sig_pool *tcp_get_md5sig_pool(void)
3175{
3176 local_bh_disable();
3177
3178 if (tcp_md5sig_pool_populated) {
3179 /* coupled with smp_wmb() in __tcp_alloc_md5sig_pool() */
3180 smp_rmb();
3181 return this_cpu_ptr(&tcp_md5sig_pool);
3182 }
3183 local_bh_enable();
3184 return NULL;
3185}
3186EXPORT_SYMBOL(tcp_get_md5sig_pool);
3187
3188int tcp_md5_hash_skb_data(struct tcp_md5sig_pool *hp,
3189 const struct sk_buff *skb, unsigned int header_len)
3190{
3191 struct scatterlist sg;
3192 const struct tcphdr *tp = tcp_hdr(skb);
3193 struct ahash_request *req = hp->md5_req;
3194 unsigned int i;
3195 const unsigned int head_data_len = skb_headlen(skb) > header_len ?
3196 skb_headlen(skb) - header_len : 0;
3197 const struct skb_shared_info *shi = skb_shinfo(skb);
3198 struct sk_buff *frag_iter;
3199
3200 sg_init_table(&sg, 1);
3201
3202 sg_set_buf(&sg, ((u8 *) tp) + header_len, head_data_len);
3203 ahash_request_set_crypt(req, &sg, NULL, head_data_len);
3204 if (crypto_ahash_update(req))
3205 return 1;
3206
3207 for (i = 0; i < shi->nr_frags; ++i) {
3208 const struct skb_frag_struct *f = &shi->frags[i];
3209 unsigned int offset = f->page_offset;
3210 struct page *page = skb_frag_page(f) + (offset >> PAGE_SHIFT);
3211
3212 sg_set_page(&sg, page, skb_frag_size(f),
3213 offset_in_page(offset));
3214 ahash_request_set_crypt(req, &sg, NULL, skb_frag_size(f));
3215 if (crypto_ahash_update(req))
3216 return 1;
3217 }
3218
3219 skb_walk_frags(skb, frag_iter)
3220 if (tcp_md5_hash_skb_data(hp, frag_iter, 0))
3221 return 1;
3222
3223 return 0;
3224}
3225EXPORT_SYMBOL(tcp_md5_hash_skb_data);
3226
3227int tcp_md5_hash_key(struct tcp_md5sig_pool *hp, const struct tcp_md5sig_key *key)
3228{
3229 struct scatterlist sg;
3230
3231 sg_init_one(&sg, key->key, key->keylen);
3232 ahash_request_set_crypt(hp->md5_req, &sg, NULL, key->keylen);
3233 return crypto_ahash_update(hp->md5_req);
3234}
3235EXPORT_SYMBOL(tcp_md5_hash_key);
3236
3237#endif
3238
3239void tcp_done(struct sock *sk)
3240{
3241 struct request_sock *req = tcp_sk(sk)->fastopen_rsk;
3242
3243 if (sk->sk_state == TCP_SYN_SENT || sk->sk_state == TCP_SYN_RECV)
3244 TCP_INC_STATS(sock_net(sk), TCP_MIB_ATTEMPTFAILS);
3245
3246 tcp_set_state(sk, TCP_CLOSE);
3247 tcp_clear_xmit_timers(sk);
3248 if (req)
3249 reqsk_fastopen_remove(sk, req, false);
3250
3251 sk->sk_shutdown = SHUTDOWN_MASK;
3252
3253 if (!sock_flag(sk, SOCK_DEAD))
3254 sk->sk_state_change(sk);
3255 else
3256 inet_csk_destroy_sock(sk);
3257}
3258EXPORT_SYMBOL_GPL(tcp_done);
3259
3260int tcp_abort(struct sock *sk, int err)
3261{
3262 if (!sk_fullsock(sk)) {
3263 if (sk->sk_state == TCP_NEW_SYN_RECV) {
3264 struct request_sock *req = inet_reqsk(sk);
3265
3266 local_bh_disable();
3267 inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
3268 req);
3269 local_bh_enable();
3270 return 0;
3271 }
3272 return -EOPNOTSUPP;
3273 }
3274
3275 /* Don't race with userspace socket closes such as tcp_close. */
3276 lock_sock(sk);
3277
3278 if (sk->sk_state == TCP_LISTEN) {
3279 tcp_set_state(sk, TCP_CLOSE);
3280 inet_csk_listen_stop(sk);
3281 }
3282
3283 /* Don't race with BH socket closes such as inet_csk_listen_stop. */
3284 local_bh_disable();
3285 bh_lock_sock(sk);
3286
3287 if (!sock_flag(sk, SOCK_DEAD)) {
3288 sk->sk_err = err;
3289 /* This barrier is coupled with smp_rmb() in tcp_poll() */
3290 smp_wmb();
3291 sk->sk_error_report(sk);
3292 if (tcp_need_reset(sk->sk_state))
3293 tcp_send_active_reset(sk, GFP_ATOMIC);
3294 tcp_done(sk);
3295 }
3296
3297 bh_unlock_sock(sk);
3298 local_bh_enable();
3299 release_sock(sk);
3300 return 0;
3301}
3302EXPORT_SYMBOL_GPL(tcp_abort);
3303
3304extern struct tcp_congestion_ops tcp_reno;
3305
3306static __initdata unsigned long thash_entries;
3307static int __init set_thash_entries(char *str)
3308{
3309 ssize_t ret;
3310
3311 if (!str)
3312 return 0;
3313
3314 ret = kstrtoul(str, 0, &thash_entries);
3315 if (ret)
3316 return 0;
3317
3318 return 1;
3319}
3320__setup("thash_entries=", set_thash_entries);
3321
3322static void __init tcp_init_mem(void)
3323{
3324 unsigned long limit = nr_free_buffer_pages() / 16;
3325
3326 limit = max(limit, 128UL);
3327 sysctl_tcp_mem[0] = limit / 4 * 3; /* 4.68 % */
3328 sysctl_tcp_mem[1] = limit; /* 6.25 % */
3329 sysctl_tcp_mem[2] = sysctl_tcp_mem[0] * 2; /* 9.37 % */
3330}
3331
3332void __init tcp_init(void)
3333{
3334 int max_rshare, max_wshare, cnt;
3335 unsigned long limit;
3336 unsigned int i;
3337
3338 BUILD_BUG_ON(sizeof(struct tcp_skb_cb) >
3339 FIELD_SIZEOF(struct sk_buff, cb));
3340
3341 percpu_counter_init(&tcp_sockets_allocated, 0, GFP_KERNEL);
3342 percpu_counter_init(&tcp_orphan_count, 0, GFP_KERNEL);
3343 tcp_hashinfo.bind_bucket_cachep =
3344 kmem_cache_create("tcp_bind_bucket",
3345 sizeof(struct inet_bind_bucket), 0,
3346 SLAB_HWCACHE_ALIGN|SLAB_PANIC, NULL);
3347
3348 /* Size and allocate the main established and bind bucket
3349 * hash tables.
3350 *
3351 * The methodology is similar to that of the buffer cache.
3352 */
3353 tcp_hashinfo.ehash =
3354 alloc_large_system_hash("TCP established",
3355 sizeof(struct inet_ehash_bucket),
3356 thash_entries,
3357 17, /* one slot per 128 KB of memory */
3358 0,
3359 NULL,
3360 &tcp_hashinfo.ehash_mask,
3361 0,
3362 thash_entries ? 0 : 512 * 1024);
3363 for (i = 0; i <= tcp_hashinfo.ehash_mask; i++)
3364 INIT_HLIST_NULLS_HEAD(&tcp_hashinfo.ehash[i].chain, i);
3365
3366 if (inet_ehash_locks_alloc(&tcp_hashinfo))
3367 panic("TCP: failed to alloc ehash_locks");
3368 tcp_hashinfo.bhash =
3369 alloc_large_system_hash("TCP bind",
3370 sizeof(struct inet_bind_hashbucket),
3371 tcp_hashinfo.ehash_mask + 1,
3372 17, /* one slot per 128 KB of memory */
3373 0,
3374 &tcp_hashinfo.bhash_size,
3375 NULL,
3376 0,
3377 64 * 1024);
3378 tcp_hashinfo.bhash_size = 1U << tcp_hashinfo.bhash_size;
3379 for (i = 0; i < tcp_hashinfo.bhash_size; i++) {
3380 spin_lock_init(&tcp_hashinfo.bhash[i].lock);
3381 INIT_HLIST_HEAD(&tcp_hashinfo.bhash[i].chain);
3382 }
3383
3384
3385 cnt = tcp_hashinfo.ehash_mask + 1;
3386
3387 tcp_death_row.sysctl_max_tw_buckets = cnt / 2;
3388 sysctl_tcp_max_orphans = cnt / 2;
3389 sysctl_max_syn_backlog = max(128, cnt / 256);
3390
3391 tcp_init_mem();
3392 /* Set per-socket limits to no more than 1/128 the pressure threshold */
3393 limit = nr_free_buffer_pages() << (PAGE_SHIFT - 7);
3394 max_wshare = min(4UL*1024*1024, limit);
3395 max_rshare = min(6UL*1024*1024, limit);
3396
3397 sysctl_tcp_wmem[0] = SK_MEM_QUANTUM;
3398 sysctl_tcp_wmem[1] = 16*1024;
3399 sysctl_tcp_wmem[2] = max(64*1024, max_wshare);
3400
3401 sysctl_tcp_rmem[0] = SK_MEM_QUANTUM;
3402 sysctl_tcp_rmem[1] = 87380;
3403 sysctl_tcp_rmem[2] = max(87380, max_rshare);
3404
3405 pr_info("Hash tables configured (established %u bind %u)\n",
3406 tcp_hashinfo.ehash_mask + 1, tcp_hashinfo.bhash_size);
3407
3408 tcp_metrics_init();
3409 BUG_ON(tcp_register_congestion_control(&tcp_reno) != 0);
3410 tcp_tasklet_init();
3411}